Lifetime estimation of lithium-ion batteries for stationary energy storage systems
Transcript of Lifetime estimation of lithium-ion batteries for stationary energy storage systems
Vattenfall R&D, KTH Royal Institute of Technology
Lifetime estimation of lithium-ion batteries for stationary energy storage systems
Degree Project in Chemical Engineering, KE202X Joakim Andersson, 9310257879 2017-06-13 Supervisors: Longcheng Liu, Jinying Yan.
Abbreviations
AFM Atomic force microscopy
ANN Artificial neural network
ARIMA Autoregressive integrated moving average
BMS Battery management system
CC-CV Constant charge-constant voltage
CDKF Central difference Kalman filter
DEC Diethyl carbonate
DEKF Duel extended Kalman filter
DMC Dimethyl carbonate
DoD Depth of discharge
EC Ethylene carbonate
EKF Extended Kalman filter
EoL End of life
ESS Energy storage system
ESVEKF Enhanced state vector extended Kalman filter
EV Electric vehicle
HEV Hybrid electric vehicle
ICA Incremental capacity analysis
LCO Lithium cobalt oxide
LFP Lithium iron phosphate
Li-ion Lithium-ion
LMO Lithium manganese oxide
LTO Lithium titanate oxide
NCA Lithium nickel cobalt aluminum oxide
NMC Lithium nickel manganese cobalt oxide
OCV Open circuit voltage
P2D Pseudo-2D
PC Propylene carbonate
PDE Partial differential equation
RC Resistor-capacitor
RLS Recursive least squares
RuL Remaining useful life
RW Random walk
SEI Solid electrolyte interface
SNN Structured neural network
SoC State of charge
SOH State of health
SPKF Sigma point Kalman filter
SPM Single particle model
STEM Scanning transmission electron microscopy
SWDEKF Single weight dual extended Kalman filter
UKF Unscented Kalman filter
VC Vinylene carbonate
ABSTRACT
With the continuing transition to renewable inherently intermittent energy sources like solar- and wind
power, electrical energy storage will become progressively more important to manage energy
production and demand. A key technology in this area is Li-ion batteries. To operate these batteries
efficiently, there is a need for monitoring of the current battery state, including parameters such as
state of charge and state of health, to ensure that adequate safety and performance is maintained.
Furthermore, such monitoring is a step towards the possibility of the optimization of battery usage
such as to maximize battery lifetime and/or return on investment. Unfortunately, possible online
measurements during actual operation of a lithium-ion battery are typically limited to current, voltage
and possibly temperature, meaning that direct measurement of battery status is not feasible. To
overcome this, battery modeling and various regression methods may be used. Several of the most
common regression algorithms suggested for estimation of battery state of charge and state of health
are based on Kalman filtering. While these methods have shown great promise, there currently exist
no thorough analysis of the impact of so-called filter tuning on the effectiveness of these algorithms in
Li-ion battery monitoring applications, particularly for state of health estimation. In addition, the
effects of only adjusting the cell capacity model parameter for aging effects, a relatively common
approach in the literature, on overall state of health estimation accuracy is also in need of
investigation.
In this work, two different Kalman filtering methods intended for state of charge estimation: the
extended Kalman filter and the extended adaptive Kalman filter, as well as three intended for state of
health estimation: the dual extended Kalman filer, the enhanced state vector extended Kalman filer,
and the single weight dual extended Kalman filer, are compared from accuracy, performance, filter
tuning and practical usability standpoints. All algorithms were used with the same simple one resistor-
capacitor equivalent circuit battery model. The Li-ion battery data used for battery model development
and simulations of filtering algorithm performance was the “Randomized Battery Usage Data Set”
obtained from the NASA Prognostics Center of Excellence.
It is found that both state of charge estimators perform similarly in terms of accuracy of state of charge
estimation with regards to reference values, easily outperforming the common Coulomb counting
approach in terms of precision, robustness and flexibility. The adaptive filter, while computationally
more demanding, required less tuning of filter parameters relative to the extended Kalman filter to
achieve comparable performance and might therefore be advantageous from a robustness and
usability perspective. Amongst the state of health estimators, the enhanced state vector approach was
found to be most robust to initialization and was also least taxing computationally. The single weight
filter could be made to achieve comparable results with careful, if time consuming, filter tuning. The
full dual extended Kalman filter has the advantage of estimating not only the cell capacity but also the
internal resistance parameters. This comes at the price of slow performance and time consuming filter
tuning, involving 17 parameters. It is however shown that long-term state of health estimation is
superior using this approach, likely due to the online adjustment of internal resistance parameters.
This allows the dual extended Kalman filter to accurately estimate the SoH over a full test representing
more than a full conventional battery lifetime. The viability of only adjusting the capacity in online
monitoring approaches therefore appears questionable. Overall the importance of filter tuning is
found to be substantial, especially for cases of very uncertain starting battery states and
characteristics.
ACKNOWLEDGEMENTS
I would like to give my deepest gratitude to all those who made this thesis a reality. To my supervisors
Longcheng Liu at the department of Chemical Engineering, KTH and Jinying Yan at Vattenfall R&D for
their unrelenting support throughout the work and for the opportunity to take part in the project in
the first place. Thank you for all the insightful input and fruitful discussions.
My thanks go out to my examiner Matthäus Bäbler for all the help and assistance I ever needed.
I would also like to thank everybody at Vattenfall R&D and KTH who helped and supported me in case
of need.
Thanks to Jonas Ricknell for insightful comments on the work.
Finally I would like to thank my mother and my father for always supporting me and for providing
motivation during rough times. I could never have made it without you.
Joakim Andersson, Stockholm, June 2017.
TABLE OF CONTENTS
1 Introduction ..................................................................................................................................... 1
1.1 Background .............................................................................................................................. 2
1.1.1 Historical perspective and outlook on lithium-ion batteries .......................................... 2
1.1.2 Lithium-ion batteries in stationary applications ............................................................. 5
1.1.3 Main components of Li-ion batteries .............................................................................. 6
1.1.4 Chemistry of Li-ion batteries ........................................................................................... 7
1.1.4.1 Cathode materials ....................................................................................................... 8
1.1.4.2 Anode materials ........................................................................................................ 12
1.1.4.3 Additives .................................................................................................................... 13
1.1.5 Battery management systems ....................................................................................... 13
1.2 Lithium-ion battery aging ...................................................................................................... 14
1.2.1 Aging of electrolyte ....................................................................................................... 16
1.2.2 Aging mechanisms at anode .......................................................................................... 16
1.2.3 Aging mechanisms at cathode ....................................................................................... 19
1.2.3.1 Chemistry-specific cathode aging processes ............................................................. 20
1.2.4 Main aging impact factors ............................................................................................. 22
1.2.4.1 Calendar aging ........................................................................................................... 23
1.2.4.2 Cycle aging ................................................................................................................. 25
1.2.5 Summary of Li-ion battery aging impact factors ........................................................... 26
1.3 Lithium-ion battery modeling ................................................................................................ 28
1.3.1 Empirical models ........................................................................................................... 29
1.3.2 Electrochemical models ................................................................................................ 29
1.3.3 Equivalent circuit models .............................................................................................. 30
1.4 State of charge estimation .................................................................................................... 32
1.5 State of health and remaining useful life estimation ............................................................ 34
2 Scope and aim of thesis ................................................................................................................. 38
3 Methodology ................................................................................................................................. 39
3.1 Equivalent circuit battery model development..................................................................... 39
3.1.1 Capacity determination ................................................................................................. 40
3.1.2 Open circuit voltage – state of charge expression ........................................................ 42
3.1.3 Identification of equivalent circuit model parameters ................................................. 46
3.2 State of health estimation validation and considerations .................................................... 50
3.3 Kalman filtering ..................................................................................................................... 53
3.3.1 Extended Kalman filter .................................................................................................. 57
3.3.2 Adaptive extended Kalman filter ................................................................................... 59
3.3.3 Dual extended Kalman filter .......................................................................................... 59
3.3.4 Enhanced state vector extended Kalman filter ............................................................. 63
3.3.5 Single weight dual extended Kalman filter .................................................................... 64
4 Results and discussion ................................................................................................................... 65
4.1 Extended Kalman filter .......................................................................................................... 65
4.2 Adaptive extended Kalman filter ........................................................................................... 76
4.3 Dual extended Kalman filter .................................................................................................. 79
4.4 Enhanced state vector extended Kalman filter ..................................................................... 88
4.5 Single weight dual extended Kalman filter ............................................................................ 92
4.6 Comparison of algorithms and summary .............................................................................. 95
4.6.1 State of charge algorithms ............................................................................................ 95
4.6.2 State of health estimation algorithms ........................................................................... 96
5 Conclusions and future work ....................................................................................................... 103
6 References ................................................................................................................................... 104
7 Appendices .................................................................................................................................. 111
Appendix 1: Code for plotting and organization of battery diagnostic data ................................... 111
Appendix 2: Code for extended Kalman filter ................................................................................. 115
Appendix 3: Code for adaptive extended Kalman filter .................................................................. 117
Appendix 4: Code for dual extended Kalman filter ......................................................................... 119
Appendix 5: Code for enhanced state vector extended Kalman filter ............................................ 122
Appendix 6: Code for single weight dual extended Kalman filter ................................................... 123
List of figures
Figure 1: Price development of Li-ion batteries 2005-2030 [21]. ........................................................... 4
Figure 2: Gravimetric energy density and specific power of different available battery technologies [2].
................................................................................................................................................................. 4
Figure 3: Simple illustration of the original LiCoO2-based Li-ion cell. Electrodes, current collectors and
separator are all submerged in the electrolyte solution [36]. ................................................................ 8
Figure 4: The three crystal structures of common Li-ion battery cathode materials. Green dots
represent Li-ions or Li-ion intercalation sites [37]. ................................................................................. 9
Figure 5: Comparison of market share and size of Li-ion battery cathode materials in 1995 and 2010
[30]. ......................................................................................................................................................... 9
Figure 6: Composition dependence for performance characteristics of NMC cathodes [46]. ............. 11
Figure 7: Graphical summary of the most important Li-ion battery degradation mechanisms [68]. ... 15
Figure 8: Example of anode side reaction leading to SEI formation via formation of DMDOHC [73]. . 17
Figure 9: Main degradation mechanisms at graphite-based anodes [40]. ........................................... 18
Figure 10: Main Li-ion battery cathode aging processes [40]. .............................................................. 20
Figure 11: Mechanism of the dissolution of manganese from LMO cathode [40]. .............................. 21
Figure 12: Aging mechanisms of layered transition metal oxides [40]. ................................................ 22
Figure 13: Effect of storage temperature on aging in terms of percentage of capacity loss. All batteries
were stored at 50 % SoC during testing [85]. ........................................................................................ 23
Figure 14: Arrhenius plot for high and low temperature behavior. r is the rate of reaction [45]. ....... 23
Figure 15: Calendar aging of LTO anode compared to graphite anode. ASI=area specific impedance. The
impedance can be seen as the equivalent of resistance for alternating current [82]. ......................... 25
Figure 16: A simple equivalent circuit model consisting of a voltage source, dependent on the SoC, and
a resistor in series. ................................................................................................................................. 30
Figure 17: n resistor-capacitor equivalent circuit model. ..................................................................... 31
Figure 18: A 1 resistor-capacitor equivalent circuit model. .................................................................. 32
Figure 19: Determining RuL by extrapolating SoH measurements [64]. ............................................... 36
Figure 20: Reference discharge voltage profiles for determination of cell capacity changes over the
course of battery testing for RW9. ........................................................................................................ 40
Figure 21: Capacity degradation during testing of RW9. Note the total timespan of around half a year.
............................................................................................................................................................... 42
Figure 22: OCV-time relationship for low current discharge test of RW9. ........................................... 42
Figure 23: OCV-SoC for RW9 battery. .................................................................................................... 43
Figure 24: Comparison of polynomial and combined model for OCV-SoC curve fitting. ...................... 45
Figure 25: Fitted OCV-SoC polynomial is clearly stable while the combined model “spikes”. ............. 45
Figure 26: All pulsed discharge tests for determination of ECM parameter values for RW9. .............. 46
Figure 27: Ballpark estimation of ECM parameters from pulsed discharge. ........................................ 47
Figure 28: 1 RC ECM model fitting results for first pulsed discharge test of RW9. ............................... 48
Figure 29: Modeling error in first pulsed discharge test of RW9 using 1 RC ECM. ............................... 49
Figure 30: Difference in modelling error when not adjusting for changing capacitance. ..................... 50
Figure 31: Voltage, current and temperature over the first 100 RW steps for RW9. ........................... 52
Figure 32: Negative SoC by Coulomb counting in RW phase 10 for battery RW9. ............................... 52
Figure 33: The Kalman filtering principle [114]. .................................................................................... 54
Figure 34: The DEKF algorithm. Solid lines represent the paths of state- and weight vectors while
dashed lines are the flows of the covariance matrices [123]................................................................ 63
Figure 35: SoC estimated by EKF over first 100 RW steps of battery RW9. .......................................... 66
Figure 36: SoC estimated EKF during end of first RW phase of battery RW9. ...................................... 67
Figure 37: SoC estimation by EKF with unadjusted ECM parameters for RW phase 5, battery RW9. .. 68
Figure 38: EKF converging to correct SoC value despite incorrect starting values. .............................. 69
Figure 39: Differences in convergence rate to Coulomb counting reference value for varying starting
SoC covariance. ..................................................................................................................................... 70
Figure 40: Eventual convergence of EKF SoC estimate despite poor choice of initial SoC covariance. 70
Figure 41: EKF SoC estimation with different measurement noise covariances................................... 72
Figure 42: Overall trend of rms SoC error versus logarithm of measurement noise covariance for cases
of noisy signals and correct starting SoC for the EKF. ........................................................................... 72
Figure 43: Cell voltage predictions of the EKF for the case of a noisy voltage signal using different
measurement noise covariance matrices. ............................................................................................ 73
Figure 44: Error bounds per (20) of SoC estimate using EKF................................................................. 75
Figure 45: The stabilization of SoC covariance in the EKF, indicating estimation convergence. .......... 75
Figure 46: SoC estimation for 100 first RW steps using AEKF with Coulomb counting reference. ....... 76
Figure 47: Rapid convergence of AEKF SoC estimate for various initial process noise covariance
matrices. ................................................................................................................................................ 77
Figure 48: The quick converging of the AEKF for various starting measurement covariances. ............ 78
Figure 49: Capacity estimation over first full RW phase of RW9. ......................................................... 80
Figure 50: SoC- (left) and capacity (right) estimates by DEKF when initial guesses of both SoC and
capacity are poor for 100 first RW steps of first RW phase of RW9. .................................................... 81
Figure 51: Rapid convergence of both internal resistance parameter estimates using the DEKF despite
poor initial guesses. ............................................................................................................................... 82
Figure 52: Measured cell voltage for RW cycling using CC-CV mode charging for RW2. Note the periods
of constant voltage at 4.2 V. ................................................................................................................. 83
Figure 53: Current over time for cycling of battery RW2. ..................................................................... 83
Figure 54: SoC by Coulomb counting and DEKF for CC-CV mode charge cycling of RW2. .................... 84
Figure 55: Convergence of capacity estimates despite large disparity in starting values using DEKF and
data from battery RW2. ........................................................................................................................ 85
Figure 56: Quick convergence of capacity estimates with different initial guesses of capacity for DEKF
using battery data from RW2. ............................................................................................................... 85
Figure 57: Capacity estimation using the DEKF over the first RW phase of RW9 with the capacity process
noise covariance set too high. Internal filter parameters are like in table 12 except for capacity process
noise covariance=1∙10-8. ........................................................................................................................ 87
Figure 58: Convergence of estimates of internal resistance parameters for varying values of process
noise covariance in the DEKF. ............................................................................................................... 88
Figure 59: Capacity estimation over entire first RW phase of battery RW9 using ESVEKF. .................. 89
Figure 60: SoC estimation using the ESVEKF for a poor initial guess of cell capacity. .......................... 89
Figure 61: Capacity estimation using the ESVEKF converging to measured value independently of
starting guess. ....................................................................................................................................... 90
Figure 62: Convergence of SoC using the ESVEKF despite poor guesses of both SoC and capacity. .... 91
Figure 63: Rapid convergence of capacity estimates with three different initial guesses using ESVEKF
and data from battery RW2 .................................................................................................................. 91
Figure 64: Convergence of capacity estimate for SWDEKF despite poor initial guess. ......................... 92
Figure 65: Convergence of capacity estimates to measured value for RW9 for different starting guesses
in SWDEKF. ............................................................................................................................................ 93
Figure 66: Capacity estimates for battery RW2 using the SWDEKF for different starting guesses of
capacity.................................................................................................................................................. 94
Figure 67: Capacity estimates using SWDEKF for RW2 for high initial capacity covariance. ................ 94
Figure 68: Comparison of long-term capacity estimation of SoH algorithms. ...................................... 97
Figure 69: Principle of "accidental" convergence of SoC estimation for faulty concurrent estimation of
capacity. The red line represents an SoC estimate with a low capacity, the green line high capacity, and
the blue line is the in-between case. .................................................................................................. 101
List of tables
Table 1: Comparison of current battery technologies [4]. ...................................................................... 5
Table 2: Common components of commercial Li-ion batteries. ............................................................. 7
Table 3: Qualitative comparison of common cathode materials [47-49]. ............................................ 11
Table 4: Advantages, disadvantages and applications of common Li-ion battery cathode materials [9].
............................................................................................................................................................... 12
Table 5: Factors affecting calendar and cycle aging. ............................................................................. 26
Table 6: Impact factors and effects of aging mechanisms [68]. ............................................................ 27
Table 7: Measured cell capacities for RW9. .......................................................................................... 41
Table 8: Development of ECM parameters with aging for battery RW9. ............................................. 49
Table 9: Internal EKF parameters for the results in figures 35 and 36. ................................................. 66
Table 10: Internal filter parameters for AEKF for figure 46. ................................................................. 76
Table 11: The effect of moving estimation window size on AEKF SoC estimation accuracy. ............... 78
Table 12: Internal filter parameter values for initialization of DEKF in figure 49. ................................ 80
Table 13: Internal filter parameters of DEKF for results in figure 51. ................................................... 81
Table 14: Differences in temperature between pulsed discharge tests and RW phases for RW9. ...... 82
Table 15: Internal filter parameters of the ESVEKF for the results in figures 59 and 60. ..................... 90
Table 16: Initial ECM parameters of battery RW2. ............................................................................... 91
Table 17: Internal filter parameters for SWDEKF results in figure 64. .................................................. 93
Table 18: Internal filter parameters for quantitative comparison of EKF and AEKF algorithms. .......... 95
Table 19: Results of comparison of EKF and AEKF for SoC estimation ................................................. 95
Table 20: Internal filter parameters for comparison of long-term capacity estimation performance of
SoH algorithms. ..................................................................................................................................... 97
Table 21: Root-mean-square capacity estimation errors for results in figure 65 for investigated SoH
estimation algorithms. .......................................................................................................................... 98
Table 22: Root-mean-square error over first four measurement points of long-term capacity estimation
test. ........................................................................................................................................................ 98
Table 23: Execution times of SoH estimation algorithms for simulations of 300 RW cycles. ............. 100
Table 1: Comparison of current battery technologies [4]. ...................................................................... 5
Table 2: Common components of commercial Li-ion batteries. ............................................................. 7
Table 3: Qualitative comparison of common cathode materials [47-49]. ............................................ 11
Table 4: Advantages, disadvantages and applications of common Li-ion battery cathode materials [9].
............................................................................................................................................................... 12
Table 5: Factors affecting calendar and cycle aging. ............................................................................. 26
Table 6: Impact factors and effects of aging mechanisms [68]. ............................................................ 27
Table 7: Measured cell capacities for RW9. .......................................................................................... 41
Table 8: Development of ECM parameters with aging for battery RW9. ............................................. 49
Table 9: Internal EKF parameters for the results in figures 35 and 36. ................................................. 66
Table 10: Internal filter parameters for AEKF for figure 46. ................................................................. 76
Table 11: The effect of moving estimation window size on AEKF SoC estimation accuracy. ............... 78
Table 12: Internal filter parameter values for initialization of DEKF in figure 49. ................................ 80
Table 13: Internal filter parameters of DEKF for results in figure 51. ................................................... 81
Table 14: Differences in temperature between pulsed discharge tests and RW phases for RW9. ...... 82
Table 15: Internal filter parameters of the ESVEKF for the results in figures 59 and 60. ..................... 90
Table 16: Initial ECM parameters of battery RW2. ............................................................................... 91
Table 17: Internal filter parameters for SWDEKF results in figure 64. .................................................. 93
Table 18: Internal filter parameters for quantitative comparison of EKF and AEKF algorithms. .......... 95
Table 19: Results of comparison of EKF and AEKF for SoC estimation ................................................. 95
Table 20: Internal filter parameters for comparison of long-term capacity estimation performance of
SoH algorithms. ..................................................................................................................................... 97
Table 21: Root-mean-square capacity estimation errors for results in figure 65 for investigated SoH
estimation algorithms. .......................................................................................................................... 98
Table 22: Root-mean-square error over first four measurement points of long-term capacity estimation
test. ........................................................................................................................................................ 98
Table 23: Execution times of SoH estimation algorithms for simulations of 300 RW cycles. ............. 100
1
1 INTRODUCTION
Due to increasing concerns of climate change there currently exists a growing interest within the
energy sector to transfer from fossil fuels such as oil, gas and coal to renewable sources of energy like
solar- and wind power. While there are several upsides to this transition, such as a lowered
environmental impact and an increased energy security, there are also several challenges that still
must be faced. One such challenge is the question of electrical energy storage. Unlike traditional
energy sources, the production of renewable energy varies depending on time of day, climate and
weather. Energy storage therefore becomes critical to manage the variable demand for electricity in
an efficient manner. Simply put, when there is an excess of energy production this should be stored
and when there is more demand than production the previously stored energy should be tapped into
to reduce the strain on the energy system. There exists a multitude of suggested technologies, so-
called energy storage systems (ESSs), intended for this purpose. These can in principle be divided into
the following categories based on the form in which the energy is stored [1]:
Electrical: capacitors, super capacitors.
Mechanical: flywheels, pumped hydroelectric systems, compressed air.
Chemical: hydrogen or other chemical storage.
Thermal: hot water, molten salts.
Electrochemical: batteries.
Battery energy storage systems, the focus of this work, possess several key advantages amongst these,
including efficiency, low pollution, rapid response time, flexibility of siting and low need for
maintenance. The modular nature of battery technologies also allows for great flexibility and
adaptability [2]. Battery power represents a very attractive option for renewable energy storage, peak-
shaving during intensive grid loads and furthermore as a back-up system for controlling voltage drops
in the energy grid, phenomena which will be of increasing importance as the transition to renewable
energy continues with the associated intrinsic climate and weather dependence [3-5].
Within the segment of battery energy storage, Li-ion batteries are of special interest, and is currently
the market-leading technology [3, 6, 7]. This is partly due to their suitability for mobile applications,
especially electric vehicles (EVs) and hybrid electric vehicles (HEVs). These are quickly developing and
growing markets, which has led to an increased interest in research within the field of Li-ion battery
technology [8]. The most rapid developments have been within the areas of Li-ion battery materials,
efficiency, reliability, safety and management systems to meet customers’ demands for vehicle
performance. The requirements of batteries for electric vehicles and for stationary energy storage
overlap in several regards. Both types of applications require long-term stability, safety, low costs and
high energy density. The differences lie mainly in the emphasized type of energy density (volumetric
density for stationary power, gravimetric density for mobile applications), and demands for flexibility
of power deliverance [2]. It is therefore believed that a synergistic development is possible and highly
advantageous. This is especially evident in the possible reuse of “worn-out” EV/HEV Li-ion batteries for
stationary applications, thus allowing effective battery lifetime to be increased by as much as a factor
of two [4].
This work aims to investigate methodologies with which to estimate the degradation of Li-ion battery
performance over time, a phenomenon referred to as battery aging. To achieve this, a simple
equivalent circuit battery model (ECM) is suggested and methods of identifying said models
parameters are investigated. The suitability of the ECM is validated using real battery data. To properly
assess battery aging an extended Kalman filter (EKF) is developed for accurate state of charge (SoC)
2
estimation and compared to the adaptive extended Kalman filter (AEKF) with updating process- and
measurement noise covariance matrices. The estimation of battery state of health (SoH), here defined
as the relative loss of battery capacity, is then examined using three different extended- or dual Kalman
filters, each integrated with the previously mentioned EKF SoC estimator. First the standard dual
extended Kalman filter (DEKF), capable of also estimating the internal resistance of the cell, is
investigated. Then two modified Kalman filter methods aiming for lower computational strain are
introduced. The first of these applies the EKF methodology but uses a state vector that has been
extended to include the cell capacity directly as a state variable. The second is based on the DEKF
methodology but reduces the weight filter to only incorporate the cell capacity. All Li-ion battery SoH
estimators are finally evaluated and compared based on accuracy, computational performance, and
stability with special regards to tuning of internal filter parameters. All data used for model- and
SoC/SoH estimation algorithm validation was obtained from the NASA Prognostics Center of Excellence
[9].
The structure of the thesis will be as follows: Below in the background section necessary theory and
earlier work within the fields of Li-ion cell construction, chemistry, aging, and modelling, with special
regard to Li-ion battery SoC and SoH monitoring will be introduced and discussed. The objective and
scope of this current work will then be outlined in chapter two. Chapter three will contain a thorough
description of the applied methodology, including Li-ion battery model development and details of the
used Kalman filtering algorithms. The obtained results using said methods are presented and discussed
in the fourth chapter and, finally, conclusions are then drawn in chapter five, along with suggestions
for future work.
1.1 BACKGROUND Lithium-ion batteries have in a relatively short time span become essential technology. With new
innovations in safety, cost, and energy density, it is predicted to remain a key competitor in the battery
sector for many years to come [10]. In this background section the goal is to provide an introduction
to Li-ion battery technology at large by considering historical, economical, physical and chemical
aspects.
1.1.1 Historical perspective and outlook on lithium-ion batteries
The Li-ion battery technology know today was developed by Asahi Chemicals and Sony during the
1980s and first became commercially available in 1991. The main goal was to come up with a novel
battery technology capable of achieving increased energy densities compared to existing solutions for
the then quickly growing market of mobile applications, mainly small portable electronics such as
mobile phones, video cameras and laptop computers [11]. The idea was to use lithium due to its low
molecular weight, leading to high energy density and fast diffusion [2, 10]. The main competing
technologies at the time were nickel-cadmium (Ni-Cd) and nickel-metal hydride (Ni-MH), both lacking
in terms of gravimetric and volumetric energy density compared to the new Li-ion technology. The
memory effect (permanent loss of capacity when not fully charging the battery) is also lower for Li-ion
batteries [12]. In fact, for a long time it was believed that Li-ion batteries suffered from no memory
effect at all but recently it has been proven to occur, at least to a minimal extent [13].
One price to pay for this increased performance was the need to use organic solvents instead of an
aqueous solution as the electrolyte for these batteries. The cells also must be pressurized. This causes
some concerns regarding safety, and special requirements for the construction of the cells. Li-ion
batteries can, when abused or improperly assembled or designed, infamously explode. One recent
example is the widely published Samsung Note scandal. Due to risks like these, testing protocols for Li-
3
ion batteries must be especially rigorous, which naturally is an added cost over other more inherently
safe battery technologies [14].
Another persisting issue has been the relatively high manufacturing cost of Li-ion batteries, mainly due
to the comparatively expensive transition metal based cathode materials [15]. Long-term there also
exists some concern regarding the use of lithium metal in the manufacturing process considering the
quite limited world inventory. Some work has therefore gone into trying to produce batteries based
on Na-ion intercalation compounds due the similarity in chemistry, sodium being an alkali metal like
lithium, as the price and the worldwide available supply of sodium is overall much healthier than for
lithium [2]. Some authors have however claimed that the price of lithium is in fact not a major driver
of battery prices, or will at least not be so long term [16].
The development of Li-ion battery technology and its market share has been drastic since its
introduction in 1991. Between 1995 and 2005 prices halved and energy density doubled. Today prices
are one tenth of what they were in 1991 and sales have increased dramatically as well. In 2013 five
billion Li-ion cells were sold just for powering of portable electronics. There is however a growing
concern within the industry that Li-ion battery performance will soon reach its peak after continual
improvements for over 25 years. Researchers currently believe that the limit is a further increase of
gravimetric density by approximately 30 percent [17].
Almost all Li-ion batteries have historically been manufactured in Japan, Korea or China, with very little
production in Europe or North America. This is also the case currently. As EVs and HEVs viable and
widely available during the early 2010s these also became a key usage areas for Li-ion batteries.
Electric- and hybrid vehicles are of course quickly developing sectors with some forecasts predicting
that 50 percent of all sold cars will use a battery for propulsion by 2020 [4, 15]. Currently Ni-MH
batteries are dominating this market, but this is expected to change rapidly. It is projected that Li-ion
batteries will be the leading format by as soon as 2020 [18]. Note however that when one only
considers fully electric vehicles, such as the Tesla Model S and X, BMW i3, Nissan Leaf and Chevrolet
Spark, Li-ion technology is already the front running technology [10, 19, 20]. It is predicted that the
increased production of Li-ion batteries for these automotive applications, along with the growing
portable electronics market, will drive down prices to the point where it becomes profitable to utilize
these in stationary battery ESS applications as well, where lead-acid batteries are currently dominating
[2, 21].
Presently Li-ion batteries are four to eight times more expensive than equivalent lead-acid alternatives
and one to four times more expensive than Ni-MH [4]. It is required that the average price of Li-ion
batteries drops further to around $125/kWh, the goal set for 2022 by the American Department of
Energy, down from $190/kWh in 2016, in order to achieve market-wide penetration in the stationary
application market [16]. Prices of first generation Li-ion batteries are expected to keep on decreasing,
despite the already dramatic lowering these past years, with an estimated 60 to 70 percent total or
approximately 7 percent annual decrease in cost per watt-hour between 2007 and 2014, with
purchasing costs for EV manufacturers dropping even more [16, 22]. Nykvist and Nilsson project a cost
reduction rate of around the same magnitude going forward as well, noting also that there currently
exists a large price discrepancy amongst suppliers. This discrepancy is also projected to decrease per
figure 1 below.
4
Figure 1: Price development of Li-ion batteries 2005-2030 [22].
For Li-ion battery powered electric vehicles to become cost competitive with traditional internal
combustion varieties it is commonly agreed that prices of Li-ion batteries need to fall to around
$150/kWh. This means that the paradigm shift to Li-ion technology is expected to occur within the
transport sector first, followed by the stationary applications at a later stage [22]. Synergistic aspects
exist between these at first seemingly quite different applications however, as shall be discussed in the
section below.
Gravimetric energy densities and specific power of different battery technologies are compared in
figure 2. The current state of the most common battery technologies is summarized in table 1.
Figure 2: Gravimetric energy density and specific power of different available battery technologies [2].
5
As can be seen in figure 2, Li-ion technology has the advantage in terms of both gravimetric- and
volumetric energy density over other common battery types. Also notable is the large difference
energy density within the category of Li-ion batteries itself. As shall be seen, Li-ion battery technology
encompasses a diverse group of chemistries.
Table 1: Comparison of current battery technologies [4].
Lead-acid Ni-Cd Ni-MH LiCoO2
(Li-ion) LiMn2O4
(Li-ion) LiFePO4
(Li-ion)
Gravimetric energy density (Wh/kg)
30-50 45-80 60-120 150-190 100-135 90-120
Cycle life (ΔSoC = 80 % each cycle)
200-300 1000 300-500 500-1000
500-1000
1000-2000
Fast charge time (h) 8-16 1 2-4 2-4 1 or less 1 or less
Self-discharge per month at room temperature
5 % 20 % 30 % <10 % <10 % <10 %
Nominal cell voltage (V) 2 1.2 1.2 3.6 3.8 3.3
Maintenance requirements 3-6 months
30-60 days
60-90 days
Not required
Not required
Not required
Toxicity High High Low Low Low Low
Peak C-rate (Ah/h) 5 C 20 C 5 C >3 C >30 C >30 C
As can be gleaned from table 1 and figure 2, the main advantages of Li-ion battery technology over its
common counterparts are its long cycle life, high nominal voltage, low need for maintenance and high
energy density. In the coming sections the usage of Li-ion batteries in stationary applications will be
considered in more detail.
1.1.2 Lithium-ion batteries in stationary applications
Usage of batteries for storage of most importantly renewable energy from solar- and wind power has
been investigated for several years. The main competing technology has generally been pumped
hydroelectric energy storage, which currently makes up 99% of all worldwide electrical grid storage
capacity [21, 23]. The main driving force in the integration of ESSs in the electrical grid has mostly been
government regulations to incentivize an accelerated transition to renewable energy generation [10].
Stationary battery applications can be split into two broad categories: energy applications and power
applications. The main difference between these is the C-rate1 of discharge, energy applications having
discharge times in hours and power applications in the range of seconds to minutes. Energy
applications such as peak shaving and load leveling involve reducing the maximum grid load at, or close
to, maximum energy demand. Grid frequency regulation and power quality control are typical modes
of operation within the power applications category [3, 21]. It is widely accepted that no currently
existing ESS is perfect in all regards. Because of this a mixture of storage systems is expected to be the
optimal solution going forward. Li-ion battery technology is expected to play a key role, particularly for
short- and medium-term storage solutions. Smaller scale applications are considered especially
suitable for Li-ion battery ESSs due to the high both volumetric and gravimetric energy density [24].
1 Charge/discharge rate is often written as C-rate, given as the amount of times the battery could theoretically be fully charged or discharged during one hour at the chosen current.
6
As mentioned above, currently around 99 % of the worldwide installed EES consists of pumped
hydroelectric systems. It is however expected that Li-ion battery technology will soon reach a point
where cost is not inhibiting its use for these purposes, and several demonstration projects are already
running with more planned [2].
As mentioned, one of the most attractive aspects of Li-ion batteries compared to other electrochemical
ESSs is its growing use for EV applications. Not only is this driving the technological development, with
an expected yearly increase in gravimetric energy density between 2 and 6 percent [19], but it is also
driving down manufacturing costs through the economy-of-scale principle. Furthermore, the prospect
of re-using batteries that have lost 20 percent of their maximum capacity through aging during
automotive applications, and would otherwise be considered ready for recycling, for stationary ESS is
promising [2, 25]. The potential residual lifetime for stationary ESS might be as long as ten years,
representing a further loss of around 15 percent capacity. This not only dramatically lowers the
lifecycle cost of Li-ion batteries as a whole, including purchasing cost for both the EV and ESS
customers, but also lowers their overall environmental impact [25]. In fact, it is currently not
considered profitable to recycle worn-out EV batteries [26]. The synergistic aspects between the
applications are also obvious: EVs could be charged using 2nd generation reused Li-ion batteries which
were in turn charged with renewable energy. The needed battery conversion to allow for stationary
use would be rather simple. Spent EV batteries would have to be tested for capacity and safety aspects
such as leakage and then repackaged with appropriate software and hardware. The main applications
for these second generation Li-ion batteries are expected to be transmission support, area regulation,
light commercial load following, load leveling, power reliability, residential load following, distributed
node telecom backup and renewable energy firming [25]. These applications can essentially be divided
into two different classes. Either large numbers of battery packs are assembled together for large scale
applications, such as integration with renewable energy generation, or smaller numbers of packs are
used for decentralized peak-shaving purposes in businesses and homes [27].
The so-called “smart grid” applications are of particular interest for second generation Li-ion batteries.
This is partly due to the expected-to-be limited supply of these second-generation batteries
considering the size of large scale ESS solutions. Smaller, distributed, decentralized ESSs for homes and
businesses can help improve efficiency and flexibility by allowing customers to regulate their personal
energy consumption. It is also considered to be safer from a financial standpoint to invest in a larger
number of smaller units than vice versa [27].
Having discussed the current and historical state of Li-ion battery markets, the next sections aim to
cover the basics of Li-ion battery construction and chemistry.
1.1.3 Main components of Li-ion batteries
The currently available Li-ion battery electrodes are based on Li-ion intercalation compounds. The goal
is for these intercalation/de-intercalation processes to be as reversible and rapid as possible to ensure
a long battery life and to have as high cell voltage as possible to increase the energy density and
achievable power from the cell. This means that demands on material performance are very high. The
first Li-ion batteries developed by Sony during the 1980’s were based on a graphite anode with a
lithium cobalt oxide (LiCoO2) cathode. Even though several improvements have been made to the
technology since then, the basic principle and materials used still largely remain the same, the cathode
material often being some transition metal oxide and the anode almost always some graphite-based
compound [28]. The most diversified aspect of Li-ion batteries is easily the cathode material with many
commercialized alternatives, each with its own strengths and weaknesses. Li-ion battery chemistries
are therefore often categorized based on the cathode material. Li-ion batteries are manufactured in
four different geometrical shapes: coin, cylindrical, flat and prismatic [29].
7
Lithium, being an alkali metal, is very sensitive to water and therefore carefully dried aprotic organic
solvent electrolytes must be used to avoid excessive lithium consumption. These are usually mixtures
of alkyl carbonate esters such as diethyl carbonate (DEC), dimethyl carbonate (DMC), propylene
carbonate (PC) and/or ethylene carbonate (EC) with a lithium salt such as lithium hexafluorophosphate
(LiPF6), lithium perchlorate (LiClO4) or lithium tetrafluoroborate (LiBF4) [30]. LiPF6 is the most commonly
found amongst these due to safety and stability reasons [31]. The lithium salt is responsible for ionic
transport through the separator and the electrolyte and should therefore preferably be as ionic as
possible. In addition, the added salt should be thermally stable, able to form stable passivating films
on both electrodes, and be as inert as possible towards all the components of the cell [32].
The solvent has three main requirements: it should have a high dielectric permittivity and a low
viscosity to allow for rapid transport of Li-ions and it should be as inert as possible to reduce lithium
consumption, thus allowing for longer battery lifetimes [33]. The goal when developing a solvent
mixture to be used as an electrolyte is to combine the advantages of the individual compounds to
achieve a good compromise of properties. It is often so that low viscosity solvents are mixed with ones
with high ionic conductivity for instance. This is as a rule done empirically [30]. Cyclic carbonate esters,
such as EC and PC, usually have the role of contributing a high conductivity while linear carbonate
esters, mainly DEC and DMC, generally constitute the lower viscosity compounds [31, 32, 34].
There will also always be a set of additives added to the electrolyte. These are introduced for several
reasons, for instance to minimize electrolyte decomposition or to improve safety. As an example,
vinylene carbonate (VC) is regularly added to improve high temperature performance [35].
A polymeric separator, usually made of polypropylene or polyethylene, prevents transport of electrons
through the electrolyte but still allows for Li-ion passage. The most popular Li-ion battery component
materials are summarized in table 2 [6]:
Table 2: Common components of commercial Li-ion batteries.
Component Materials Characteristics
Cathode LiCoO2 (LCO), LiNi1-x-yMnxCoyO2 (NMC), LiMn2O4 (LMO), LiFePO4 (LFP), LiNi1-x-yCoxAlyO2 (NCA)
High voltage
Anode Carbonaceous material (natural/artificial graphite, hard/soft carbon)
Low voltage
Electrolyte Non-aqueous solvents + Li salt + additives (DEC/DMC/PC/EC + LiPF6/LiClO4/LiBF4 + VC)
Li-ion conductor
Separator Porous polymer membrane (PP, PE) Electrical insulator
In addition, current collectors, most often made of copper on the anode side and aluminum on the
cathode side, help with the transport of electrons to/from the cell on charging/discharging.
In large scale applications, such as stationary energy storage or in EVs and HEVs, a single battery system
may consist of thousands of connected Li-ion cells. The cells can be connected in series or in parallel
configurations. More cells will be connected in parallel if a higher current is required and more in series
for a higher voltage. Connecting the cells in parallel also allows for a higher total capacity [36].
1.1.4 Chemistry of Li-ion batteries
The basic chemical processes occurring in the Li-ion battery during operation, as for any
electrochemical cell, are the redox reactions at the cathode and anode. During discharge of the
battery, oxidation is taking place at the anode and reduction at the cathode. During this process, Li-
ions are deintercalated from the anode, transported through the electrolyte, and finally intercalated
into the cathode crystal lattice. At the same time electrons are also transferred in the same direction,
8
that is from anode to cathode, but through an external circuit to the load. The discharge reaction is
spontaneous, meaning that energy can be extracted from the process. This is the basic principle which
allows for electricity to be outputted from the Li-ion cell. The process is also exothermal, meaning that
heat is produced. The driving force for the intercalation/deintercalation process is difference in
potential between the cathode and anode, which is dependent on the cell chemistry, the current state
of charge, and operation conditions but normally lies between 3 and 4 V for Li-ion cells. Upon charging
the direction of this process is reversed and the anode is re-stocked with intercalated lithium ions. Due
to this back-and-forth motion of the lithium ions on charging and discharging these batteries are
frequently referred to as “rocking chair batteries”. For a generic metal oxide LiMO2 cathode cell the
partial redox reactions are the following on discharge:
Anode: 𝑥𝐿𝑖𝐶6 ⇒ 𝑥𝐿𝑖+ + 𝑥𝑒− + 𝑥𝐶6 (A)
Cathode: 𝑥𝐿𝑖+ + 𝑥𝑒− + 𝐿𝑖1−𝑥𝑀𝑂2 ⇒ 𝐿𝑖𝑀𝑂2 (B)
The most common alternatives for the various Li-ion components will now be discussed, starting with
cathode materials followed by anode materials, and then finally the important electrolyte additives. A
simple sketch of a Li-ion cell illustrating its working principle is shown below in figure 3.
Figure 3: Simple illustration of the original LiCoO2-based Li-ion cell. Electrodes, current collectors and separator are all submerged in the electrolyte solution [37].
1.1.4.1 Cathode materials
There currently exists a wide range of commercialized alternatives for Li-ion battery cathode materials.
Generally speaking, the cathode material should be able to achieve a high energy density and have a
high electric conductivity. A high energy density is effectively equivalent to being able to intercalate a
large number of Li-ions. In addition, Li-ions should diffuse quickly through the material and the material
should be as inert as possible towards common electrolytes to reduce aging effects [34]. The most
common cathode materials can be divided into three groups based on their crystal structure. Lithiated
layered transition metal oxides such as LiCoO2 (LCO), LiNiO2 (LNO), LiNi1-x-yCoxAlyO2 (NCA) and LiNi1-x-
yMnxCoyO2 (NMC); olivine structured lithium iron phosphate (LiFePO4 (LFP)) and spinel structured
lithium manganese oxide (LiMn2O4 (LMO)) [12]. There currently does not exist a universally optimal
cathode material. All the discussed materials have distinct advantages and disadvantages as well as
9
key commercial applications for which they are most suitable. The three common cathode material
crystal structures are seen below in figure 4.
Figure 4: The three crystal structures of common Li-ion battery cathode materials. Green dots represent Li-ions or Li-ion intercalation sites [38].
The three structure types are sometimes referred to as 3D (olivine), 2D (layered), or 1D (spinel) based
on the dimensionality of Li-ion motion through the materials [39]. Although LCO remains the most
prominent cathode material overall, its market share has been decreasing continuously over time,
mostly due to cost and safety aspects. A comparison of the cathode material market in 1995 and in
2010 is shown in figure 5 below. Note also the massive growth of the total market in that time: from
650 tons of cathode materials in 1995 to 45,000 tons in 2010, almost a 70-fold increase.
Figure 5: Comparison of market share and size of Li-ion battery cathode materials in 1995 and 2010 [31].
A more detailed look at the various common cathode materials will be given in the coming sections.
1.1.4.1.1 Lithium cobalt oxide
The first commercialized Li-ion batteries in 1991 utilized the LCO cathode. LCO remains the most
common Li-ion battery cathode material currently, even though it has been losing ground to other
chemistries. Cobalt metal itself is quite expensive, heavy, and the LCO based batteries have poorer
thermal stability than most other commercialized alternatives. Aging in terms of capacity loss is fast
during high current-, deep discharge-, cycling which might limit suitability for some applications [40].
Its most popular use is in portable energy devices due to the high C-rate capability and the high
volumetric energy density [3]. Safety issues remain a concern for LCO-based batteries. Recently many
incidents have been reported for two relatively new applications of LCO-based batteries: hoverboards
and e-cigarettes, where the batteries would self-ignite, once again causing the inherent safety of these
types of batteries to be put into question [10].
10
1.1.4.1.2 Lithium nickel oxide
LNO is very similar to LCO in terms of its physical structure and theoretical capacity. It was originally
introduced as a cheaper alternative but still possesses a higher energy density (20 percent higher by
weight). LNO is unfortunately quite difficult and expensive to produce in a pure enough form as Ni2+
ions have a tendency to take the place of Li+ ions during synthesis and cycling [40], especially at high
temperatures, causing the reduction of Ni3+ ions for charge balancing in the process [41]. LNO also
displays issues of structural Jahn-Teller distortion, safety concerns and exothermic release of oxygen
at higher temperatures [39]. The material is therefore not commonly used [3]. It was however a key
development on the way to NCA- and NMC based cells [32, 39].
1.1.4.1.3 Lithium nickel cobalt aluminum oxide
NCA is a modified version of LNO that incorporates aluminum to produce a more stable material while
still maintaining most of the advantages of LNO. The capacity of NCA is not quite as high as for LNO,
mainly due to the fact that the Al3+ ions do not partake in the electrochemistry of the cell, but its aging
behavior is instead superior which has led to widespread usage [32]. Unfortunately, these batteries
are especially sensitive to moisture during their assembly, leading to increased manufacturing costs
and the need for strict environmental control. The upside is that these cells are suitable for both high
energy and high power applications and are generally considered to be of high quality in terms of
performance. The batteries supplied by Panasonic for Tesla EVs make use of this cathode material for
instance [40].
1.1.4.1.4 Lithium manganese oxide
LMO was the second cathode material to be commercialized after LCO but had already been suggested
by Thackeray et al. as early as 1983 [31, 42]. Manganese is approximately five times cheaper than
cobalt and is also quite plentiful in nature by comparison. Manganese based materials also have the
advantage of being more environmentally benign than cobalt based ones. LMO has a higher nominal
voltage than LCO- or LNO-based chemistries. The energy density is however around 20 percent lower
by weight than LCO and cycle stability is relatively poor. The fast battery aging on cycling is usually
attributed to the dissolution of Mn2+ into the electrolyte. This mechanism will be investigated further
in section 1.2.3.1.1. On the other hand, LMO-based batteries have higher thermal stability and they
are cheaper as they do not require expensive cobalt or nickel [3]. This battery chemistry is often used
for applications requiring high power outputs, such as power tools, due to its high voltage [32]. Another
promising route is to blend in LMO with NMC cathodes to reduce costs, increase capacity, and improve
thermal stability [32].
1.1.4.1.5 Lithium iron phosphate
Another commercialized alternative is LFP which was first introduced in 1997 [43]. The main
advantages of the iron phosphate-based electrode materials are the flat voltage profile, reduced cost,
increased stability against common electrolytes, excellent high temperature performance, and
practically non-existent toxicity. Electron conductivity, energy density and theoretical capacity are
however lower than for other common alternatives. The lower potential does have the added benefit
of reducing issues of electrolyte oxidation and the resulting battery aging thus improving lifetime and
safety [3]. Some authors consider this to be the most promising cathode material for large scale energy
storage for these reasons, especially when combined with a titanate anode to further reduce
degradation rate and increase safety [2, 4, 32, 39, 44]. The issues of low ionic and electronic
conductivity could be resolved by addition of dopants, metal dispersions or nanostructuring with the
downside of increased manufacturing cost. Nanostructuring would also likely result in a lowered
volumetric energy density [39, 45].
11
1.1.4.1.6 Lithium nickel manganese cobalt oxide
On the other end of the voltage spectrum from iron phosphates we find the so-called high voltage
electrode materials. These materials do not have any advantages over conventional cathode
alternatives in terms of capacity, but instead focus on achieving a high discharge voltage (close to 5 V
is possible). One of these is NMC which is a promising replacement material for the traditional LCO due
to increased cycling and thermal stability as well as higher capacity [39]. The ratio of
nickel/manganese/cobalt in NMC batteries varies and can be adjusted as to fit the intended
application. Roughly speaking, increasing relative manganese content increases safety, increasing
relative nickel content improves capacity, and increasing relative cobalt content results in higher
achievable C-rates [46]. This is summarized in figure 6 below. Recently NMC has been used extensively
for automotive applications and is currently the cathode chemistry of choice for EVs made by BMW,
Volkswagen, Fiat, Kia and Ford [10]. A newer development of NMC, LiNi1/2Mn3/2O4 belongs to the
category of high voltage electrode materials but possesses a spinel structure much like LMO. Synthesis
of these materials is challenging however and they are quite unstable with regards to the commonly
used electrolytes at high voltages [3, 32].
Figure 6: Composition dependence for performance characteristics of NMC cathodes [47].
A simple comparison and summary of common cathode materials is presented in the table below [36].
Table 3: Qualitative comparison of common cathode materials [48-50].
LCO NMC NCA LMO LFP
Cell voltage 3.7-3.9 V 3.8-4.0 V 3.65 V 4.0 V 3.3 V
Energy ++ +++ +++ + ++
Power ++ ++ +++ +++ ++
Calendar life + + +++ - ++
Cycle life + ++ ++ ++ ++
Safety + + + ++ +++
Price estimate -- ++ + ++ +
Due to the various strengths and weaknesses of the Li-ion battery cathode materials, the optimal
choice of material is largely application-dependent, as seen in table 4 below [3, 36].
12
Table 4: Advantages, disadvantages and applications of common Li-ion battery cathode materials [10].
LCO NMC NCA LMO LFP
Advantages Cycle life, energy density
Excellent energy density
Cycle life, power Thermal stability, price, energy density
Excellent cycle life
Disadvantages Thermal stability
Patent issues Sensitive to moisture
Cycle life Energy density, power
Common applications
Portable electronics
Power tools, EVs
High quality electronics
Power tools Power tools, stationary energy storage, e-bikes
1.1.4.2 Anode materials
Due to the relatively low capacity and stability of the commonly used carbon based electrode
materials, researchers have been trying to find viable alternatives. These are all still lithium
intercalation compounds. It has however seemed to be difficult to find materials that can compete
with carbon in terms of overall compromise between performance and cost. The diversity of Li-ion
battery anode materials is therefore much lower than that of cathode materials. Titanate, as
mentioned below, is the only commercialized alternative [36, 51, 52].
Besides the advantage of low cost, carbon based materials are also abundantly available, display a low
potential vs. Li/Li+, allow for fast Li+ diffusion, and show relatively little volume change on Li-ion
intercalation/deintercalation, leading to low mechanical stress on cycling. Common carbon anodes are
either based on graphite (soft carbon) or hard carbon. Graphite based materials have the advantage
of high capacity, but lack compatibility with PC based electrolytes due to excessive side reactions. On
the other hand, hard carbons generally have a lower capacity but are instead more stable over time
with regards to aging [40].
Tin oxide anodes of composition SnMxOy (M=B, P, Al) were first suggested in 1997. The main advantage
that these materials provide over traditional graphite is a higher energy density. Tin oxide anode
materials will however degrade relatively quickly on cycling and due to the persistence of this issue
they have not been commercialized as of yet. The practical energy density is also frequently
substantially lower than predicted theoretical values. Composites of tin-oxide with various other
compounds have therefore been investigated to alleviate these issues, for instance with carbon
nanotubes, magnesium oxide, silicon oxide, zinc oxide, or cobalt oxide. This is very much an active
research area [51].
The fact that LTO could be used as a material for reversible Li-ion intercalation/deintercalation was
first reported in the 1980s [53]. LTO is unique as an anode material in that its potential is essentially
independent of the degree of intercalation of Li-ions. This flat voltage curve allows for nearly constant
power delivery over the entire SoC range of the battery. Also, the change in volume on intercalation
of Li-ions is practically non-existent. This reduces aging due to mechanical stress dramatically. For
traditional graphite electrode materials, the volume change on intercalation is around 10 percent
which, long-term, leads to fissures, cracking and exfoliation of the anode material. These processes are
investigated closer in section 1.2.2. The great disadvantages of LTO are the low electron conductivity
along with the rather extensive release of gas by-products, predominately H2, CO and CO2, during
operation at high temperatures [54]. These issues have limited the use of LTO cells for large-scale
applications [55]. Researchers have been trying to improve the material properties in these regards
through doping or protective coatings [51]. In addition, LTO-based batteries have lower energy
13
densities and capacities than graphite based ones due to the higher weight and also higher potential
vs. Li/Li+ (about 1.5 V higher than graphite anodes) [56].
Silicon-based anode materials have also been suggested due to their potential very high capacity for
intercalation of Li-ions. In fact, Silicon has the highest theoretical potential for Li-ion intercalation of
all commonly found elements, around 4200 Ah/kg. The most common alternative is thin-film
electrodes of amorphous silicon. Historically, the issue with these materials has been the large volume
change on intercalation/de-intercalation, causing cracking and even pulverization of the material.
Electron conductivity of Silicon is also generally quite poor. Widespread commercialization of Si-based
anode materials has therefore not taken place [51, 57].
1.1.4.3 Additives
Additives are always added to Li-ion battery electrolytes with the general intentions of reducing aging
effects and improving performance. More specifically, the main goals are to improve the anode SEI
stability, to reduce LiPF6 decomposition, to protect the cathode, to reduce or stabilize metallic Li
deposition and to improve overall safety. Certain additives might also be added to reduce the corrosion
of the aluminum current collector, to improve the wetting of the polymer separator, or to act as a fire
retardant [58]. The most commonly added additive, as mentioned above, is VC. The purpose of VC is
to limit the loss of active lithium due to SEI formation by forming a thin film on the anode. VC is reduced
and polymerizes before SEI formation, thus improving battery life by avoiding a loss active cycling Li-
ions. Other additives, such as lithium bis(oxalato)borate (LiBOB) will react with already formed SEI
compounds to produce more stable ones, thus effectively suppressing further SEI growth and
subsequent decomposition [32].
One of the major sources of side reactions in common Li-ion batteries leading to loss of capacity and
increased internal resistance is the decomposition of the common LiPF6 salt to form hydrofluoric acid
(HF) via a series of reactions. This is especially evident at higher temperatures and needs to be
suppressed as much as possible to ensure long term battery performance. The first step in this process
is the following reaction:
𝐿𝑖𝑃𝐹6 ⟺ 𝐿𝑖𝐹 + 𝑃𝐹5 (C)
PF5 is the compound that will go on the produce not only HF but also CO2 and the toxic POF3. The last
two compounds are both in gas form and their formation will therefore lead to a pressure build-up in
the cell, a serious safety concern. This equilibrium can be shifted towards the reactant side by simply
adding LiF to the electrolyte which is frequently done. Another strategy is to reduce the reactivity of
already formed PF5 by adding complexing agents, often tris(2,2,2-trifluoroethyl) phosphite. Finally, HF
or H2O scavengers, such as butyl amine, are also frequently added as a last resort solution [32, 58]. The
mechanisms involved in HF formation and the related Li-ion battery aging processes will be explored
further in section 1.2.
1.1.5 Battery management systems
The specialized combination of hardware and software for real-time, online monitoring and control of
battery performance is called the battery management system (BMS). A BMS is necessary for all Li-ion
batteries due the high demands for strict control of operation parameters to ensure safe and efficient
performance [59-61]. The first BMSs were developed in the 1990s by Texas Instruments, Ericsson and
Motorola [8]. The tasks of the BMS can be divided into two broad categories [61, 62]:
14
Optimization of operating parameters such as C-rate, ΔSoC, and cell voltage to ensure safe,
efficient performance. This means keeping these parameters within certain battery-specific
operation windows, often referred to as the safe operating area (SOA). These parameters also
should be balanced evenly among all individual cells in the battery pack. This concept is called
equalization.
Continuous monitoring of vital battery parameters to determine battery SoC, SoH, remaining
useful life (RuL) etc. via algorithms. This is called battery monitoring.
These two categories are of course closely interconnected. Operation outside of the SOA induces
accelerated aging and can potentially damage the battery irreversibly. For instance, if the operation
temperature drops below the recommended limit, lithium plating may deposit on the anode resulting
in a loss of capacity and available power. Accurate estimation of say the SoH is likewise critical in
determining the appropriate SOA for the optimization procedure as the optimal location of this
window will certainly change as the battery ages [63].
The general issue of SoC-, SoH- and RuL determination is that the internal chemical processes of the
Li-ion battery are not, at least directly, observable during operation. This means that one must identify
critical variables that can be measured during operation that can somehow be used to indicate SoC,
SoH and/or RuL via some functional relationship. It is often the case that the measurable parameters
are limited to just current, voltage, and temperature [64]. Algorithms for online monitoring of SoC and
SoH will be discussed in sections 1.4 and 1.5 respectively. It is often the case that the limited
computational power and memory of the BMS are key factors to consider when developing a battery
monitoring method, as is the accuracy of available measurement sensors. Measuring voltage with
reasonable accuracy is most often not an issue. Accurate current measurement can however prove to
be a challenge. In addition, the battery behavior will change over time due to various aging
mechanisms, thus further complicating the work of the BMS. For these reasons the algorithms for SoC-
, SoH-, and RuL estimation must strike a balance between accuracy, numerical robustness, flexibility
and computational performance, which is not an easy task [62, 65]. To identify and indicate the extent
of battery aging, one must first recognize the causes and symptoms of the involved physical and
chemical processes, which is not elementary and understanding of these processes remains an active
research area [66]. The most important aspects of Li-ion battery aging will be introduced in the
subsequent section.
1.2 LITHIUM-ION BATTERY AGING For the ideal battery, the redox reactions at the cathode and anode are completely reversible and the
only chemical reactions occurring within the cell. For a real battery, this is not the case. The Li-ion
battery will lose capacity and gain internal resistance over time due to various side reactions. These
performance degradation processes are collectively referred to as battery aging. Battery aging reduces
available power and energy of the battery and eventually leads to the battery having to be
decommissioned, often as a safety precaution [67, 68].
Understanding of the chemical and physical processes that lead to battery aging is critical in optimizing
battery usage, lifetime, chemistry and construction [69]. The issue is very challenging due to the large
number of possible reactions and the interactions of these, along with the highly limited possibilities
of collecting data during operation of the battery. As mentioned, most often only voltage, current and
temperature can be measured during online operation. These problems become even more complex
when one considers equalization aspects for entire battery modules and/or packs, consisting of several
cells connected in series and/or parallel [63]. This work will focus on only the cell level aging
phenomena.
15
The exact nature of the battery aging processes depends on current and historic operation conditions
as well as the cell chemistry. While the differences in aging behavior between different cell chemistries,
as will be discussed further in coming sections, can be quite substantial, some general trends of all Li-
ion battery aging can be elucidated. The most important processes in Li-ion battery aging are [67]:
Formation, decomposition and precipitation of a solid electrolyte interface (SEI) on the anode.
Deposition at anode (dendrite formation, lithium plating).
Metal dissolution from cathode.
Loss of active electrode material (particle cracking, structural disordering, dissolution,
exfoliation).
The growth of a passivating layer on the anode, commonly referred to as the SEI, is of particular
interest. Many researchers consider this to be the most important aging process at the anode of Li-ion
batteries, to the point where this is commonly the only mechanism considered in physics-based aging
models [69]. The carbon anode is in fact not thermodynamically stable with respect to the common
organic solvent electrolytes and chemical reactions between the electrolyte and the carbon will always
occur to some extent. This is however not an issue during normal operation. The SEI actually protects
the anode from corrosion and from degradation reactions with the electrolyte, essentially acting as a
safety barrier [69, 70]. Figure 7 summarizes the most common aging effects in Li-ion batteries.
Figure 7: Graphical summary of the most important Li-ion battery degradation mechanisms [69].
As can be seen in the figure above, the number of possible degradation reactions is very large.
Determining the exact nature of these reactions, and furthermore their interactions, is often
exceedingly difficult. It is also clear that several length scales must be considered simultaneously to
get the complete picture of Li-ion battery aging; from the molecular to the macroscopic [107]. Of
course, several time scales must be considered as well, from the near-instantaneous Li-ion
intercalation/de-intercalation reactions to the slow growth of the SEI over the operating time of,
potentially, months and years. This separation of time scales makes accurate modeling very
challenging, especially considering the inherently slow rate of aging data acquisition, such as capacity
fade.
In the coming sections component-specific aging phenomena of common electrolyte and electrode
materials will be summarized and discussed.
16
1.2.1 Aging of electrolyte
While the electrolyte does not provide any chemical function in the battery bar ionic transport, it does
play a key role in battery aging. Reactions between the electrolyte and the electrodes particularly so,
as will be discussed in subsequent sections. It is however also important to consider the processes
occurring within the confines of the electrolyte itself as these will often be the starting point of other
reactions at the electrodes, the separator and/or current collectors.
Perhaps the most important of these electrolyte-internal processes is the formation of HF. After PF5
has been formed via the decomposition of LiPF6, see reaction (C) above, it can react further with traces
of water, simple alcohols or other protic contaminants in the electrolyte to form HF and phosphoryl
fluoride (POF3) [41, 71]. Here the reaction with water is shown.
𝑃𝐹5 + 𝐻2𝑂 ⇒ 𝑃𝑂𝐹3 + 2 𝐻𝐹 (D)
It is thought that the formation of PF5 is accelerated via interactions with the common electrolyte
constituents EC and DEC. These reactions are more prominent at elevated temperatures. Formed HF
may react with Li-ions to form LiF, releasing a proton in the process and thus lowering the pH of the
electrolyte. This will in turn lead to further HF formation by the following LiPF6 interactions [71, 72].
𝐻+ + 𝑃𝐹6− ⇔ 𝐻𝐹𝑃𝐹5 (E)
𝐻𝐹𝑃𝐹5 ⇒ 𝐻𝐹 + 𝑃𝐹5 (F)
The formation of HF may therefore be seen as autocatalytic. Formed PF5 will also react with the
electrolyte, particularly EC. This will cause polymerization of EC to occur under release of CO2. The
interactions of the Li-salt, most commonly LiPF6, and the electrolyte are generally quite complex and
not yet fully understood [71]. The most common strategies for dealing with HF formation are addition
of additives, as discussed in section 1.1.4.3, and drying of solvents to very low moisture content to
avoid reaction (D) from occurring in the first place. Avoiding water is overall considered a key aspect
in limiting Li-ion battery aging. Li-ion cells are therefore normally assembled at low air humidity and
electrodes are dried at elevated temperatures [73].
1.2.2 Aging mechanisms at anode
According to many researchers the most important aging processes in Li-ion batteries occur at the
anode, at least for cells with graphite based anode materials. As these types of anodes are by far the
most common, the present discussion of anode aging mechanisms will be focused on mainly these
materials. Again, formation of the SEI is commonly considered the most important process. The
formation of the SEI occurs right from the start of battery operation. The carbon anode will
immediately upon the first charging of the battery react with EC, DMC, DEC, and/or EMC in the
electrolyte to form various solid phase compounds such as ROCO2Li and CO2OLi, consuming active Li-
ions in the process and thus lowering the available cell capacity. These formed compounds can then
react further with traces of carbon dioxide or water in the electrolyte to form compounds like
dimethyl-2,5-dioxahexane dicarboxylate (DMDOHC), ethylmethyl-2,5-dioxahexanedicarboxylate
(EMDOHC) or diethyl-2,5-dioxahexane dicarboxylate (DECDOHC) [68, 74]. An example reaction is
shown in figure 8, a suggested route to DMDOHC. Note that CO2 is formed during the reactions pictured
in figure 8, which might cause further side reactions or cell deformation due to the increased internal
pressure.
17
Figure 8: Example of anode side reaction leading to SEI formation via formation of DMDOHC [74].
In addition to these types of reactions, the LiPF6 salt dissolved in the electrolyte may react with traces
of water to form PF5 and HF as discussed above. These react with Li-ions to form other solid, insulating
compounds, LiF for example, that accumulate on the anode as part of the SEI as well. The traces of HF
may also cause LIFMn3+ in Mn-containing cathodes, mainly LMO, to disproportionate into Mn4+ and
Mn2+. This will cause dissolution of cations from the cathode in the organic electrolyte. These cations
will then migrate to the anode under charging conditions and form insoluble depositions on its surface,
adding to the SEI [58, 68].
As mentioned above, the existence and slow growth of the SEI will not be a major issue at normal
battery operating conditions, even though it will reduce the number of active cycling Li-ions, and thus
available cell capacity. Furthermore, SEI growth will also slowly increase cell internal resistance over
time due to the lower Li-ion diffusion rate in the growing surface layer. SEI formation is expected and
can thus be accounted for. The concerning aging processes in the SEI occur at non-optimal
temperatures or during overcharging of the cell. At high temperatures, the SEI will break down and
crack. The SEI might also crack due to overcharging or storage at high SoC for prolonged periods. The
cracking of the SEI will expose bare graphite surface which leads to accelerated formation of new SEI,
thus reducing the cell capacity as active Li-ions are consumed in the process. The reason for the higher
SEI formation rate is believed to be the catalytic effect of the exposed fresh graphite surface [58]. SEI
cracking might also cause loss of electrical contact and loss of available active sites in the anode for Li-
ion intercalation due to blockage. This might also cause graphite exfoliation. The risk of SEI cracking
increases with layer thickness, meaning that these processes become increasingly important at the
later stages of the battery life [68, 75].
The structural changes of the graphite itself during cycling will also cause aging. The intercalation and
de-intercalation of Li-ions, particularly at higher C-rates and ΔSoC, along with solvent co-intercalation,
causes the graphite layers to become increasingly disordered due to fissures, splits and cracking of the
carbon material [76]. Chemical reactions producing gas might also cause cracking, especially during
storage at high SoC for prolonged periods [70]. The loss of order in the graphite generally decreases
electrode Li-ion capacity due to a loss of active sites. In addition, loss of order or cracking of the
graphite anode will generally lead to increased SEI formation due to the increased exposed area of
fresh graphite, thus leading to increased loss of active Li-ions as described above [77].
18
Beside SEI formation and growth and the loss of structural order within the electrode, lithium plating
is yet another important graphite anode aging process to consider [67, 69]. Li-ions will under certain
circumstances not be intercalated into the anode but will instead be reduced to form a metallic lithium
deposition on its surface, thus adding on to the SEI [78]. This effect is especially evident when:
The anode capacity is lower than the cathode capacity, for instance due to anode aging;
The temperature is low;
The C-rate under charging is high.
This may be rationalized by imagining an accumulation of Li-ions on the surface of the anode due to
slow intercalation kinetics (low temperature) or lack of intercalation sites (low electrode capacity). The
competing lithium plating reaction will then dominate due to the large Li-ion concentration gradient
at the surface of the electrode [78]. The composition of the electrolyte and electrode materials will
also affect the extent and rate of lithium plating [79]. Lithium plating causes a loss of active cycling Li-
ions and an increase of internal resistance due to growth of the passivating SEI on the anode in the
process. It is interesting to note that lithium plating seems unique among the Li-ion battery aging
processes in that its rate is accelerated at lower temperatures [80]. Dendritic growth of metallic lithium
depositions on the anode is a serious safety hazard. These dendrites might grow to breach the
separator and thus cause catastrophic cell failure in the form of short-circuiting, possibly leading to
ignition or explosion [37, 80, 81]. Due to the anode capacity dependence of the lithium plating
reactions, the effect will usually be more evident at the later stages of battery life when anode capacity
has been reduced and intercalation kinetics have slowed down due to various aging processes. Lithium
plating will however tend to slow down after its onset as the resulting change in capacity balance
between anode and cathode will counteract the effect [33, 80]. Figure 9 below summarizes the main
degradation mechanisms at graphitic anodes.
Figure 9: Main degradation mechanisms at graphite-based anodes [41].
19
Lastly, it is worth mentioning that the main aging mechanisms are quite different for LTO-based
anodes. First and foremost, no SEI formation will occur on LTO anodes due to their higher potential vs.
Li/Li+ compared to graphitic anodes. For the same reason, no lithium or manganese plating will occur
either. As mentioned in section 1.1.4.2, the volume change of LTO upon Li-ion intercalation is minimal
which means that mechanical stress on cycling is also lessened. The main concern of LTO anode aging
is instead the quite substantial production of gas at high temperatures. Still, it is commonly so that
aging processes occurring at the cathode will dominate the overall aging behavior of cells with an LTO
anode [53, 82, 83].
1.2.3 Aging mechanisms at cathode
In this section the characteristic aging mechanisms of different cathode materials will be discussed.
While it generally is the case that processes as the anode will dominate the overall aging behavior of
the cell, again, especially for the common graphite anodes, the cathode might also play an important
role, depending on operation conditions and cell component materials. In particular, cathode aging
may become the dominating factor for operation under especially harsh conditions, such as elevated
temperatures [33]. First general aging mechanisms involving the cathode will be summarized, followed
by a more detailed look at the specific aging behavior of common cathode materials.
The three main types of degradation processes are the same for all cathode materials: structural
changes, metal dissolution and the formation of a passivating film. Sometimes this cathode film is also
referred to as a SEI. One then must distinguish the anode SEI from the cathode SEI. When one talks
about simply “the SEI” one will usually mean the anode SEI.
Like for the graphitic anode, volume changes caused by intercalation/de-intercalation of Li-ions during
charging/discharging will induce structural changes in the cathode. This causes mechanical stress and
may lead to phase transitions occurring in the material. These phase transitions may then cause further
mechanical stress due to differences in specific volume between the phases. Phase transitions will also
lead to a reduction of electrode capacity due to a loss of available sites for Li-ion intercalation. For
these reasons, cathode materials will usually be doped in such a way as to reduce volume changes
upon lithiation, for instance by incorporating aluminum or magnesium into layer structured cathode
materials and cobalt, chromium, aluminum or magnesium into spinel structured ones [41].
Again, just like for the common graphite anode, there will also be film formation on the common
cathodes due to side reactions with electrolyte. These reactions involving the electrolyte will however
generally be slower and affect aging less than the ones occurring at the graphitic anodes. The cathode
film will increase cell internal resistance and thus the available power of the cell. The cause of these
reactions is the oxidation of the electrolyte by the electrode materials and/or LiPF6 decomposition [70].
Under normal operating conditions aging at the anode will dominate. As temperature increases
however, the electrolyte oxidation is believed to become the crucial aging process. The main Li-ion
battery cathode aging mechanisms, common to all cathode materials, are summarized in figure 10
below.
20
Figure 10: Main Li-ion battery cathode aging processes [41].
Due to the large differences in structure and chemistry of the common cathode materials, these will
also generally display different aging behavior, at least in terms of what processes are the most central
to the overall electrode aging. The most important differences between the aging processes of the
common cathode materials will now be outlined. In terms of aging behavior, the cathode materials are
divided into three categories, based on crystal structure: LMO, representing spinel structure materials;
LFP, representing olivine structured materials; and the layered transition metal oxides, represented by
LCO, LNO, NMC, and NCA.
1.2.3.1 Chemistry-specific cathode aging processes
1.2.3.1.1 Lithium manganese oxide
Due to the unique spinel structure of LMO and redox behavior of manganese, the dominating aging
processes for these cathode materials are quite different to those occurring for the layered- and olivine
structured cathode compounds. Manganese can exist in three different oxidation states: +2, +3 or +4.
In the Li-ion battery cathode LiMn2O4, manganese is normally present in the oxidation states +3 or +4.
When discharging an LMO-based battery at high C-rates, especially to low cell voltages, extra Li-ions
may be intercalated into the cathode. This causes a reduction of Mn-ions from Mn+4 to Mn3+ in the
cathode. The cause for this over-intercalation of Li-ions is the much slower diffusion of these in the
cathode material compared to in the electrolyte. This causes an accumulation of Li-ions around the
LMO cathode particles during discharge. This also allows for rationalization of the acceleration of this
process during conditions of high C-rates and low voltages. A high C-rate indicates that more Li-ions
are transferring in the cell and low voltage means fewer available sites for Li-ion intercalation in the
cathode. This leads to an increased concentration gradient of Li-ions at the electrolyte-electrode
interface, leading to increased rate of over-intercalation. The resulting reduction of Mn+4 to Mn3+
induces Jahn-Teller distortion, leading to a phase transition and corresponding volume increase of the
cathode material. This phase transition also leads to a loss of cathode capacity due to a loss of active
sites for Li-ion intercalation. The formed Mn+3 may then disproportionate into Mn4+ and Mn2+. The
21
formed Mn+2 dissolves in the organic electrolyte, as discussed in section 1.2.2. This process is catalyzed
by traces of HF or other acids in the electrolyte [39, 56]. The effect is also more prominent at higher
temperatures, partly due to the faster kinetics of the reaction itself and partly due to the increased
rate of HF formation [32, 41]. The dissolution of manganese might contribute to as much as 20 to over
30 percent of the total capacity loss of LMO cathodes [39].
Due to these reduction and disproportionation processes, unique to these types of cathodes, LMO-
based materials are especially sensitive to high C-rates and low voltages. To alleviate this problem
somewhat, researchers have tried to develop special inorganic coatings of Al2O3, ZrO2, MgO or AlF3 to
combat over-intercalation, or special manganese trapping separators to stop the migration of already
dissolved Mn2+ to the anode [84]. The manganese dissolution process is illustrated in figure 11.
Figure 11: Mechanism of the dissolution of manganese from LMO cathode [41].
Looking at figure 11 it is again seen how important it is to limit the exposure of the cell to water.
1.2.3.1.2 Lithium iron phosphate
Due to the stronger chemical bonds to the oxygen atoms, LFP materials are generally more stable at
higher temperatures compared to both the layered transition metal oxides and LMO, in the sense that
fewer side reactions with the electrolyte will occur. In fact, cell voltage would have to be as high as 5.4
V for oxidation of common electrolytes to be at all possible, far higher than the normal operation
window of LFP cells [39]. There does exist an HF catalyzed iron dissolution mechanism somewhat like
the dissolution of manganese from LMO cathodes, but it is considered to be of less concern for the
overall aging. Dissolution of iron results in a loss of active material and thus battery capacity along with
an increased internal resistance due to the growth of the anode SEI [44, 56]. In addition, the lower cell
voltage when using these cathode materials leads to a lowered reactivity towards the electrolyte
during storage [32]. Due to the low ionic and electronic conductivity of LFP cathodes they are especially
sensitive to high C-rates, which will lead to accelerated aging [39].
22
1.2.3.1.3 Layered transition metal oxides
The dissolution of metal ions is in general not a concern for these compounds except for during
operation at overly high voltages (beyond 4.2 V vs. Li/Li+). Note that this goes even for the manganese-
containing NMC as manganese is only present in the inactive +4 form in this material. Capacity fade
due to phase transitions is also less of a concern [73]. The formation of a cathode SEI is accelerated by
high voltages and, in addition, high temperatures. Even though these processes will occur for all
cathode materials, it is of special importance for this class of electrodes. These transition metal oxides
can act as oxygen donors in themselves by forming a phase of low lithium/oxygen with a rock-salt
structure. This rock-salt phase is also a detriment to Li-ion transport through the electrode, being of
lower conductivity and thus increasing cell internal resistance. The electrolyte is oxidized during this
process, and transesterification or polymerization reactions may occur, depending on the solvent
composition and conditions [41]. During these reactions gas may be released, potentially causing
pressure build-up and subsequent deformation of the cell [85]. While all layered transition metal
oxides share this behavior, it is especially prevalent for LCO, and the less common LNO, cathodes. As
the Li-ion content in these cathodes drops below around 50 percent of the maximum intercalated
amount, electrolyte oxidation becomes severely limiting and for this reason one should be careful not
to operate batteries with a LCO or LNO cathode at low SoCs frequently [39]. The most important aging
processes of layered transition metal oxides are summarized in figure 12 below.
Figure 12: Aging mechanisms of layered transition metal oxides [41].
1.2.4 Main aging impact factors
This section aims to summarize the various factors that affect battery aging divided into calendar- and
cycle aging as is often done in the literature. Calendar aging is defined as the aging of the battery that
is occurring while it is not being operated, that is when no current is entering or leaving the battery.
Cycle aging is the aging happening when the battery is being either charged or discharged.The goal is
to recapitulate the most commonly stated functional relationships between operating conditions and
aging rate. This is essential for an overall understanding of the degradation processes and their
interconnections. It is also the first step to developing a holistic aging model and to optimize battery
operation. As shall be seen, it will not be necessary to differentiate between various aging mechanisms
in battery monitoring per se, as the actual effects on battery performance largely overlap. This allows
for the definition of a relatively simple metric of overall battery aging status, the state of health.
23
1.2.4.1 Calendar aging
The main factors affecting the extent and rate of calendar aging are cell chemistry, temperature,
calendar time and the SoC during storage. High temperatures will accelerate the rate of all side
reactions, resulting in losses of active electrode material as well as of active cycling Li-ions. The effect
increases exponentially as seen in figure 13, thus suggesting a Arrhenius-like description [86].
Figure 13: Effect of storage temperature on aging in terms of percentage of capacity loss. All batteries were stored at 50 % SoC during testing [86].
Very low temperatures will on the other hand lead to lithium plating on the anode becoming the
dominating aging process [33]. This alludes to two different aging regions based on temperature: at
low temperatures lithium plating on anode will dominate, and at high temperatures SEI formation on
both electrodes (more so for the layered metal oxides, less so for LMO and especially LFP), as well as
dissolution of manganese for LMO based cells, will govern the overall aging [46]. Because of this
acceleration of aging at both high and low temperatures it is quite difficult to encompass the total
behavior in one Arrhenius-like expression for all temperatures [87]. For operation within a limited
range of temperatures this might be sufficient however [88, 89]. The temperature dependence of
calendar aging has been reported to be different for different cell chemistries and even cell
geometries. Waldmann et al. found that lithium plating will dominate under 25 °C, while SEI formation
is more important above 25 °C for the investigated NMC cells and therefore suggested using two
different Arrhenius-like expressions for these ranges, see figure 14 below [46].
Figure 14: Arrhenius plot for high and low temperature behavior. r is the rate of reaction [46].
24
Bandhauer et al. reached a similar conclusion, stating that battery aging will be very rapid over 50 °C,
especially so for storage at high SoC due to the growth of the SEI reducing available amounts of Li-ions,
and under -10 °C due to lithium plating on charging. Because of the risk of dendritic growths of lithium
on the anode, charging at high C-rates when temperature is low should be avoided as far as possible
as this might cause short-circuiting which might be a serious safety risk. Thermal monitoring and
control is therefore essential for Li-ion batteries, particularly so for larger applications [90].
As for the effect of SoC on calendar aging, a higher SoC during storage implies a state further from the
thermodynamic equilibrium of the cell. This in turn makes reactions between electrodes and
electrolyte more favored. The exact mathematical relation is however not clear. Exponential [91], but
also linear [88], SoC-dependencies of capacity loss during storage have been reported. Su et al. noted
two distinctly different aging behaviors below and above 95 percent SoC during storage and attributed
this to a change in the dominating aging mechanism perhaps due to different compounds being formed
and incorporated into the SEI at SoCs above 95 percent [92]. One way of introducing the SoC
dependence of calendar aging is to proceed via the Tafel equation which describes the rate of an
electrochemical reaction as a function of the overpotential, the overpotential being closely related to
the storage SoC [70].
Total calendar aging, measured in either capacity loss or increase of internal resistance, will not be
linear over time but will generally be more rapid during the earlier stages of the battery’s life [70, 93].
Several authors have concluded a time dependence of t1/2 for calendar aging [70, 86, 89, 94-96]. This
has been widely attributed to the continuous loss of Li-ions due to the SEI formation on the anode [97].
Indeed, first-principles kinetic investigations suggest that this should be the case for graphite anodes.
However, larger exponents such as 0.75 have also been reported. It is believed that this might be an
indication of a more complex relationship between calendar aging and storage time, meaning that
other mechanisms than just SEI formation are contributing [81, 88]. As SEI formation on the anode
appears to at least be the dominating aging mechanism during storage, cells with titanate-based
anodes should see relatively little calendar aging as no SEI is formed on these electrodes.
The calendar aging behavior of Li-ion batteries, measured as capacity fade, appears to be a memoryless
or path independent process, meaning that future aging only depends on the current state of the
battery and future operating conditions. It may therefore be considered a Markovian process. This
also indicates that capacity loss represents a good measure of overall battery aging as the history of
capacity fade does not need to be considered to predict its future development [89, 92]. The
importance of this fact will become evident in the upcoming sections on battery modeling and capacity
fade estimation.
Eddahech et al. investigated the role of battery chemistry on calendar aging. NMC, NCA, LFP and a
LMO-NMC hybrid were compared. These were stored at a set of conditions of SoC and temperature. It
was determined that the LMO- and NMC-based cathodes were particularly sensitive to high storage
temperatures. Dissolution of manganese was confirmed to be a leading factor in the aging for these
cells by post-mortem analysis. LFP was most resistant to degradation at all temperatures while NCA
was found to be the middle-of-the-road choice [98].
The work done on calendar aging of cells featuring LTO-based anodes suggests that these materials
will vastly outperform those with graphite based anodes in terms of capacity and internal resistance
retention over time. Belharouak et al. compared the calendar aging of two LMO-cathode cells with all
things the same except for the anode material, one being traditional graphite and one being LTO. Both
cells were held at 70 percent SoC and 55 ⁰C during the 28-day test, except for when measurements of
internal resistance using a current pulse method were performed once per day. The striking results are
25
shown in figure 15 [83].
Figure 15: Calendar aging of LTO anode compared to graphite anode. ASI=area specific impedance. The impedance can be seen as the equivalent of resistance for alternating current [83].
As can be seen in figure 15, the calendar aging of the LTO-based cell, measured in increase of area
specific impedance, is almost non-existent over the course of testing, especially compared to the
graphite-based cell. The authors attribute this extreme difference to the lack of SEI-formation in the
LTO cell.
1.2.4.2 Cycle aging
Cycle aging is occurring when the battery is either being charged or discharged. Due to the dynamic
nature of the processes occurring in the battery during cycling, the associated aging phenomena are
more complex and multi-faceted. The number of impacting parameters is also larger. Of course, it is
also impossible to separate out the always ongoing effects of calendar aging during tests of cycle aging,
further complicating the study of these processes [70].
One important factor to consider is the variation of SoC, or ΔSoC, during cycling, which is usually given
as an average value. It has been demonstrated that aging will be accelerated for high ΔSoC cycling. This
might be related to the mechanical stress induced by the large change in intercalated lithium in the
electrodes and the accompanying volume change, particularly for the cathode [99].
The C-rate appears to also play a key role. A high C-rate will, due to exothermic side reactions and the
internal resistance of the cell leading to Joule heating, increase the working temperature, thus further
increasing the rate of side reactions. Some authors have therefore included thermal effects into their
aging models to account for this. The most straightforward method of accomplishing this is to include
only Joule heating due to internal resistance, which is easily calculated from an equivalent circuit model
for instance [91, 100]. The higher rate of intercalation/de-intercalation of Li-ions resulting from the
higher C-rate also implies accelerated transport of Li-ions through the cell. At very high C-rates there
is a chance that Li-ions are not completely intercalated/de-intercalated from the electrodes, leading
to increased rates of lithium plating and dendrite formation [64, 67]. In addition, synergistic effects
may exist. For instance, a low temperature along with a high C-rate and low voltage will accelerate
lithium plating further [41].
26
Interestingly, Wang et al. introduced a C-rate dependence of cycle aging by noting that the activation
energy term in their developed Arrhenius-like aging expression became smaller at higher C-rates,
thereby including a factor like the one below in the final expression describing capacity loss to
accommodate for this effect [97].
𝐶𝑙𝑜𝑠𝑠 ∝ (𝑒
𝐴+𝐵∗𝐶𝑟𝑎𝑡𝑒𝑅𝑇 )
Where Closs is the absolute loss of capacity compared to the new battery, R the gas constant, and T the
temperature in Kelvin. A and B are constants, where A is large and negative and B positive. The model
equation was found to fit experimental data well at varying C-rates.
Schmalstieg et al. investigated the influence of not only ΔSoC, but also of the average SoC during
operation on cycle aging. Interestingly, it was noted that battery capacity degradation was accelerated
at both high and low average SoC cycling, with a minimum at around 50 percent average SoC. This then
means that if one wants to minimize the effects of the always occurring calendar aging by keeping the
average SoC low, the rate of aging due to cycling will increase instead. Therefore there should exist an
optimal average SoC where overall aging is minimized for a given battery usage pattern or application
[88].
As aging due to cycling accumulates over the lifetime of the battery, there needs to be some
quantitative measure of how much cycling has occurred to properly account for this effect. The two
most popular options for lifetime models are cycle number counting and calculation of cumulative Ah-
throughput. The Ah-throughput is the total transferred charge during the lifetime of the battery [41,
101]. Often some exponential relationship between Ah-throughput and capacity loss is found to
appropriately fit data [88, 91]. When using cycle number to track the cumulative transferred charge it
is necessary to define what is exactly meant by a cycle. Li-ion batteries will typically not be just
repeatedly fully charged and discharged after all. One method of transferring current-time data to
equivalent cycles is via the rainflow-counting algorithm as done in [96, 102].
1.2.5 Summary of Li-ion battery aging impact factors
As can be concluded from the discussion above, Li-ion battery aging is a highly complex issue. This is
partly due to the number of influential factors but also the interactions of these. Furthermore, it is
clear that the area of application and the usage pattern are essentially important to the aging process
through the dependencies on C-rate, average SoC and ΔSoC. The main factors affecting calendar and
cycle aging discussed in section 1.2 are summarized in table 5 below.
Table 5: Factors affecting calendar and cycle aging.
Impact factors for calendar aging Impact factors for cycle aging
Cell geometry
Cell chemistry
Temperature
SoC during storage ΔSoC during cycling
Calendar time C-rate
Cycle number or Ah-throughput
Average SoC during cycling
Factors placed in the middle of table 5 affect both calendar- and cycle aging. Effects of calendar and
cycle aging are normally considered additive. This means that the full aging behavior can be described
using one modeling function with some factors being only dependent on average SoC, temperature
and time, these thus effectively representing calendar aging. Whether this is representative of the
27
fundamental physical- and chemical processes causing Li-ion battery aging is very difficult to determine
due to the interconnectedness of the mechanisms and their effects [86, 91, 103].
The aging mechanisms discussed above will generally lead to a loss of battery capacity and/or an
increased internal resistance. The effects of Li-ion battery aging, common mechanisms, and associated
important impact factors are summarized in table 6 below, based on the discussions in previous
sections [69].
Table 6: Impact factors and effects of aging mechanisms [69].
Mechanisms Effects Impact factors
SEI growth Capacity loss, resistance increase
Time, high temperature, high voltage, high C-rate
SEI decomposition, cracking
Capacity loss High temperature, high voltage, high C-rate
Electrolyte oxidation Capacity loss High temperature, high voltage
Graphite exfoliation Capacity loss, resistance increase
High voltage, high C-rate
Lithium plating Capacity loss, resistance increase
Low temperature, low voltage
Electrode cracking Capacity loss, resistance increase
Mechanical stress, high C-rate
Transition metal dissolution (Mn, Fe)
Capacity loss, resistance increase
High C-rate, low voltage, high temperature
Here it is also evident why the degradation of cell capacity is a sensible choice for an indicator of overall
battery health when discussing battery aging. As is evident in table 6, all major aging mechanisms lead
to a loss of battery capacity while only some lead to an increase of internal resistance. Therefore,
capacity loss may be considered a better representation of the overall battery health in the general
sense. Of course, also being able to also estimate the increase of internal resistance allows for a more
nuanced picture of the occurring aging-related chemical processes.
It has also been seen that aging characteristics are dependent on cell chemistry, especially so for
different anode materials. LTO-based cells display superior calendar- and cycle aging compared to
those featuring a graphite anode. In addition, the main aging mechanisms are very different for LTO-
based cells as no SEI formation and hardly any electrolyte oxidation will occur in cells with these
anodes. Most physics-/chemistry based aging models of Li-ion battery aging are based on cells with a
graphite anode and these will therefore generally not describe the aging in LTO-based cells accurately.
It is also interesting to note that overall aging behavior may depend on the interconnection of cells in
parallel and/or series. Minimizing variations in thermal conditions, SoC and capacity loss between cells
in the same battery pack might be crucial in preventing accelerated aging due to overcharging.
Fortunately perhaps, thermal and electrical balancing are somewhat interconnected in that hotter cells
are also capable of faster C-rates due to the lowered internal resistance as a result of the increased
temperature [90]. Obviously, for cells connected in series the cell with the lowest capacity determines
the capacity of that whole system [91]. Even though these effects seem important, relatively little
material has been published on the aging behavior of entire packs or modules of cells, perhaps partly
due to the increased need for measurement equipment and the increased complexity [63]. The optimal
BMS would estimate the capacity of all cells in the system simultaneously and then distribute load so
that the aging of cells was as even as possible, naturally an enormously challenging task [5].
28
Due to the vast number of interconnected processes that make up Li-ion battery aging, and because
all mechanisms have not yet been fully understood, it is currently not possible to apply a first-principles
aging model for battery monitoring. As an accurate model is essential to achieve acceptable results,
several simplified approaches have emerged to tackle the issue. Another reason as to why simplified
battery models need to be used is to ensure satisfactory computational performance. Due to the
transient nature of Li-ion battery behavior, it is essential that the sampling time of measurements is
sufficiently brief to allow for this behavior to be accurately captured. In this sense, a complex model
might, slightly ironically, actually reduce overall accuracy of battery state estimation as the sampling
time might need to be too long for the algorithm to “keep up” with new measurements. Reducing
model complexity also increases hopes of flexibility as to be usable for different battery chemistries,
geometries and operation conditions. This is because fewer cell characterization measurements need
to be made to parameterize the model. In fact, as shall become evident in later sections on for instance
dual extended Kalman filtering, online characterization might be an attractive option for certain
battery models.
1.3 LITHIUM-ION BATTERY MODELING In the clear majority of applications, the measurement data available for determination of battery
status during actual operation is essentially limited to battery voltage, current and (surface)
temperature. To indirectly estimate battery state and to monitor its status using just these relatively
few measurements, a battery model is necessary. For instance, determining the battery SoC and/or
SoH is often a goal of monitoring. A satisfactory battery model should be able to capture and replicate
the behavior of the real battery to an acceptable extent whilst still achieving adequate computational
performance [47]. The limiting factor for computational performance is frequently the processing
power of the BMS for online solutions, offline methods usually having more leeway in this regard. The
most commonly applied battery models can be divided into three categories:
Purely empirical models,
Electrochemical models,
Equivalent circuit models.
Most current BMSs utilize either empirical- or equivalent circuit models [5]. There also exist
crossbreeds of the three categories, such as analytical-electrochemical models [104]. Due to their
relative obscurity, these types of combined approaches will not be discussed further. Out of the three
main categories above, the electrochemical models may be considered the most advanced while also
being the most intensive computationally. The empirical models are found at the other end of the
spectrum, often utilizing relatively simple equations to simulate battery behavior and thus achieving
good computational performance, this at the cost of some accuracy. The equivalent circuit models
(ECMs) seem to currently constitute the happy medium between these two extremes, achieving both
respectable performance and accuracy. In addition, the values of the ECM parameters, at least for
simpler model variants, have physical interpretations, thus facilitating a more intuitive handling of
issues such as battery aging. The three different modeling approaches will now be discussed in more
detail, starting with empirical models.
29
1.3.1 Empirical models
Empirical models can generally be divided into analytical and stochastic models. Examples of analytical
models include the Nernst model, the combined model and the kinetic battery model (KiBaM). These
models use relatively simple equations to describe battery dynamics in a purely pragmatic sense
without any physical or chemical interpretation. The KiBaM, which was originally developed for lead-
acid batteries, will for instance model the battery charge (currently available capacity) as being in one
of two hypothetical wells, either the available charge well or the bound charge well. The flowrate
between these two wells depends on the difference in height of the two “liquid levels” in the wells
[105]. A newer extended version of the KiBaM is the spatial kinetic battery model [106].
Stochastic models will generally model battery behavior as a discrete Markovian process, meaning that
a probabilistic approach is used. Like the analytical models, these models often utilize a high level of
abstraction [106].
While the empirical models might encapsulate battery behavior well, they have the quite major
downside of their parameters lacking physical interpretation meaning that they are not particularly
well suited for estimation of SoH or RuL [104].
1.3.2 Electrochemical models
These models are based on electrochemical kinetics and mass transfer equations. The most common
approaches are the single-particle model (SPM) and the pseudo-two-dimensional model (P2D), the
SPM being more simplified and the P2D more accurate, broadly speaking. The SPM models the
electrodes as being spherical particles with the same total area as the real porous materials, with
transport phenomena inside said hypothetical particles being considered. The SPM validity is limited
for cases of high C-rates, larger cells, and when electrolyte effects are significant. The growth of the
SEI layer can easily be included into the SPM to model aging effects such as loss of capacity [5, 107].
The P2D takes more physical- and chemical phenomena into account than the SPM, such as Butler-
Volmer kinetics and Li-ion concentrations in both the liquid- and solid phases of the cell. Here the
electrodes are modelled as consisting of several identical particles [108].
These types of models are often very computationally demanding, requiring the numerical solution of
a set of interconnected PDEs in each iteration of the algorithm. This means that online monitoring is
not always feasible. The large number of fitted parameters might also be questionable and overfitting
may prove to be a real issue. In addition, more information about the battery needs to be gathered
through characterization tests to parameterize these models in comparison to ECMs. This means that
more experiments need to be performed before operation can begin, which might be expensive and
time consuming [109]. The P2D and especially the SPM are very simplified models which have been
developed to try to alleviate the computational strain of other electrochemical models. This seems to
come at the sacrifice of some accuracy however and these simplified models are still slower than most
equivalent circuit models [104].
On the other hand, electrochemical have higher potential of future development than ECMs in at least
two regards. Firstly, electrochemical models could potentially be used for predictions of battery
performance for varying geometric parameters and conditions and, secondly, side-reactions leading
to aging could be incorporated directly into electrochemical models [109].
30
1.3.3 Equivalent circuit models
It appears that ECMs constitute the most popular battery models for SoC and SoH estimation in the
literature, presumably as they seem to provide a reasonable compromise between accuracy and
computational performance. The basic principle of an ECM is to model the dynamic behavior of the
battery using simple hypothetical electric circuits consisting of elements such as resistors, voltage
sources, Warburg elements and capacitors connected in series and/or parallel configurations. These
elements may be seen to represent some real physical phenomena occurring in the battery, at least
for simpler equivalent circuits [110]. The simplest possible ECM consists of a voltage source coupled
to a resistor in series, as seen in the figure below.
Figure 16: A simple equivalent circuit model consisting of a voltage source, dependent on the SoC, and a resistor in series.
Here Vt is the measurable terminal cell voltage; OCV is the open circuit voltage of the cell, a
thermodynamic quantity specific to the cell chemistry; R represents the internal resistance of the cell,
as per Ohm’s law. The OCV of the cell is readily calculated using the equilibrium potentials of the
electrodes of the cell:
𝑂𝐶𝑉 = 𝐸𝑐𝑎𝑡ℎ𝑜𝑑𝑒 − 𝐸𝑎𝑛𝑜𝑑𝑒 (1)
The equilibrium potentials of the Li-ion battery electrodes are dependent on the stoichiometry, that is
the amount of intercalated lithium ions, or the SoC, the standard potential of the electrode materials,
and weakly on temperature. In addition, it is also affected by battery aging through changes in
electrode geometries, such as changes to crystal structure and/or morphology, and loss of Li-ions
available for cycling [109, 111]. This effect is however small and can be neglected for most purposes
as will be the case in this work [65].
While the model above is extremely simplified, it still captures some fundamental battery behavior.
For instance, the OCV will change as a function of SoC, that is depending on how much charge is left in
the battery, and the cell voltage will also drop during load due to the internal resistance of the battery
per Ohm’s law:
𝑉𝑡 = 𝑂𝐶𝑉(𝑆𝑂𝐶) − 𝐼 ∙ 𝑅 (2)
The inclusion of resistor elements into the ECM also allows for thermal modeling by utilizing Joule’s
first law:
𝑃 ∝ 𝐼2 ∙ 𝑅 (3)
Here P is the dissipated power. In a real battery, there are however additional effects which one might
want to consider for a more realistic model. Most importantly, a real cell will not respond immediately
31
to an applied current. It will also not instantly decay back to the OCV after being rested. This is
essentially due to diffusion effects inside the cell. This effect is most often modelled by a series of
parallel resistor-capacitor (RC) couples like in the ECM below:
Figure 17: n resistor-capacitor equivalent circuit model.
Where Ci is the capacitance of the i:th capacitor and Vi the transient voltage over the i:th resistor-
capacitor pair. Using Kirchhoff’s current law, which states that the sum of all currents flowing into and
out of a node in an electrical circuit must be zero, and Ohm’s law, the transient voltage over a given
resistor-capacitor pair for the case of zero current may be written as:
𝐶𝑖
𝑑𝑉𝑖
𝑑𝑡+
𝑉𝑖
𝑅𝑖= 0 (4)
This differential equation gives the following solution:
𝑉𝑖 = 𝑉𝑖,0 ∙ 𝑒−
𝑡𝑅𝑖∙𝐶𝑖 (5)
Vi,0 is the transient voltage over the i:th resistor-capacitor pair at zero time. Equation (5) clearly shows
the exponential decay of the transient voltage when current is cut off. When a current I is applied to
the sub-circuit the following differential equation is the result:
𝑑𝑉𝑖
𝑑𝑡= −
1
𝑅𝑖 ∙ 𝐶𝑖∙ 𝑉𝑖 +
1
𝐶𝑖∙ 𝐼 (6)
The differential equation (6) may be converted to discrete time which gives the following result,
without going over the details of the derivation. As shall be seen, the conversion to discrete time is a
requirement for the application of the model equations in Kalman filtering methods anyway. The
subscript k represents the time-step and Δt is the time between the time-steps k and k+1.
𝑉𝑖,𝑘+1 = 𝑒
−∆𝑡
𝑅𝑖∙𝐶𝑖 ∙ 𝑉𝑖,𝑘 + 𝑅𝑖 ∙ (1 − 𝑒−
∆𝑡𝑅𝑖∙𝐶𝑖) ∙ 𝐼𝑘
(7)
This expression has two terms. The first represents the exponential decay of the transient voltage over
time when no current is applied and the second represents the transient voltage response to a
charging- or discharging current. With this expression in hand, it is simple to add RC couples to an ECM
by just adding on more terms on the form of (7) for every RC couple. The terminal voltage for an n RC
ECM can simply be written as:
32
𝑉𝑡 = 𝑂𝐶𝑉(𝑆𝑂𝐶) − 𝐼 ∙ 𝑅0 − ∑𝑉𝑖
𝑛
𝑖=1
(8)
An expression on this form will be used in this work. Hu et al. compared a set of twelve different battery
ECMs commonly applied for Li-ion battery modelling. It was found that the 1 RC model performed the
best overall in terms of compromise between accuracy of estimation and computational performance
and this will therefore be the battery model of choice in this work as well. Interestingly both Nejad et
al. and Hu et al. also noted that the most suitable ECM was found to be somewhat dependent on
battery chemistry, with the 1 RC hysteresis model achieving better results for LFP-based cells [104,
110]. Of course, real battery behavior is much more complex with concentration- and temperature
gradients, current and material distributions just to name a few aspects which are difficult to account
for using ECMs [100]. There is however no denying the success that these models have had historically
[104]. A simple sketch of a 1 RC ECM, as will be used in this work, is shown below with accommodating
notation for all equivalent circuit elements.
Figure 18: A 1 resistor-capacitor equivalent circuit model.
The accuracy of an n RC ECM may be increased by increasing the number of RC couples or by including
some SoC- or temperature dependence into the ECM parameter values [112]. Due to the lack of
appropriate battery diagnostic data, this was not possible in this work however.
1.4 STATE OF CHARGE ESTIMATION For the discussion of battery technologies there exist a set of field-specific concepts. As the exact
characterization of these will vary slightly from author to author the definitions used in this work will
now be specified for future reference. First the state of charge (SoC) of a battery will be defined as
follows:
𝑆𝑜𝐶 =
𝑄
𝐶=
∫(𝜂 ∙ 𝐼)𝑑𝑡
𝐶 (9)
Here Q is the current remaining charge in Ah and C the current maximum capacity in Ah of the battery.
Maximum capacity is defined as the amount of charge in Ah that can be discharged from a fully charged
battery, meaning a battery at the upper cut-off voltage, until the lower cut-off voltage is reached. Note
that C will change over time as the battery performance degrades due to various aging processes.
33
Lastly, η is the Coulombic efficiency of the cell, a measurement of how close the cell gets to total charge
conservation defined as discharged/charged ampere hours. For most Li-ion cells, η is very close to one
as side reactions generally are of little concern. If the Coulombic efficiency is assumed to be exactly
equal to one, as is the case in this and many other works on Li-ion battery monitoring, it does not need
to be included in equation (8) and one can simply calculate the transferred charge by summarizing the
product of current and Δt over all time steps [82].
The SoC is a measurement of how much charge can be discharged from the battery in its present state
as a percentage of the maximum available charge. The maximum available charge is also known as the
capacity. This is then the EV equivalent of the remaining fuel gauge in internal combustion vehicles,
again with the key difference that the “fuel tank” in EVs shrinks over time [64]. The concept and
definition of SoC has long been debated. Here a simple engineering definition is employed for the sake
of convenience. One should however keep in mind that a “true” SoC must be measured or estimated
at thermodynamic equilibrium, unaffected by chemical-, thermal-, mechanical or other gradients,
which is not so easily achievable and also extremely time consuming [109].
When comparing capacities, it is important to acknowledge that the capacity is sensitive to especially
temperature (measured as being 7.7 percent higher at 45 ⁰C than at 15 ⁰C in [113]) but also the C-rate
(higher discharge capacity at lower C-rates) to some extent [113]. Both effects can be rationalized by
considering the accelerated chemical kinetics and mass transfer at elevated temperatures and C-rates.
In this work, these effects were not considered as data on temperature and C-rate dependency was
not available for the examined batteries, but it is worth noting that these effects exist and will affect
the results.
The remaining charge can be measured via simply integrating the battery current over time if the initial
remaining charge is known, for instance if the battery is initially fully charged (then Q0= C):
𝑄 = 𝑄0 − ∫𝐼𝑑𝑡
(10)
As the maximum capacity of a new battery is frequently specified by the battery supplier, the SoC can
then simply be calculated by dividing (10) by this known quantity. This method of calculating SoC is
commonly referred to as Coulomb counting. With Coulomb counting it seems as if though the issue of
calculating SoC has been solved as it simply requires measuring the applied and discharged current
to/from the battery. The method has some quite severe shortcomings however. Firstly, the initial SoC
of the battery must be known. This is as a rule only the case when the battery has been fully charged
(SoC=1) which might be too rare an occurrence for certain applications to achieve reliable, accurate
results. Secondly, due to the summation over all time steps, all slight biases and other errors of
measurement sensors will accumulate over time, thus gradually degrading estimation accuracy.
Thirdly, the maximum capacity must be known beforehand. For anything but a new battery this will
generally not be the case and the changes to the capacity must be determined experimentally or
estimated by some other method as the battery ages.
In other words, Coulomb counting is far from perfect [109]. For this reason, several other, more
advanced approaches have been developed. The starting point of most of these is to not only utilize
current data but also cell voltage for SoC estimation. This is often referred to as closed-loop SoC
estimation. Despite these innovations Coulomb counting remains the most popular SoC estimation
method in the industry [114].
34
SoC estimation is a relatively explored but still challenging field. The basic requirements of a SoC
estimator are computational performance, accuracy and robustness. The essential detail of these
algorithms is the utilization of cell voltage data along with the current data. The cell voltage behavior
can be modelled as a function of the SoC, either via the OCV or directly and can thereby act as a
calibration factor for the SoC calculated via Coulomb counting. The accuracy of SoC determination is
essentially limited by the accuracy of the cell voltage modelling. Of course, offline SoC estimation
during resting of the battery is readily performed given that the OCV-SoC relation is known. If enough
time has passed since the interruption of operation, overvoltages will have decayed fully and the
“pure” OCV is then directly measurable. This method is less suitable for cell chemistries with
particularly flat OCV-SoC curves, especially LFP, as voltage measurements must be extremely accurate
to ensure a precise SoC determination [5]. For most battery applications, this is however of little
interest as operation must be halted for extended periods to perform the measurement [65].
Therefore, only online SoC estimation will be considered further in this work.
Online SoC estimation algorithms are plenty and remains a very active field of research. See [65] and
[109] for in-depth reviews on the subject. The most popular methods are based on regression, utilizing
one of the battery models outlined in section 1.4 of this work and an algorithm for interpretation of
measured voltage, current and, possibly, temperature data. There exist many algorithms capable of
serving this purpose. The most common are based on Kalman filters such as the linear Kalman filter
(LKF) [115], the extended Kalman filter (EKF) [116], or sigma point Kalman filters (SPKFs) [117],
arranged in order of increasing capability of the Kalman filter of handling non-linear systems. A smaller
number of algorithms based on state observers such as the sliding mode observer [118] or the simpler
Luenberg observer [119] have also been suggested.
1.5 STATE OF HEALTH AND REMAINING USEFUL LIFE ESTIMATION The battery state of health (SoH) is an indicator of the extent of battery performance degradation due
to aging processes. The need for a catch-all concept is evident when one considers the immense
complexity of battery aging along with the highly limited possibilities of diagnosing the nature of the
aging processes during actual battery operation. Unfortunately, there exists no set definition of SoH
and several different metrics are used in the industry [64]. Most often SoH is defined as the maximum
capacity of the battery in its current state as a percentage of the initial nominal capacity of the fresh
battery as below. Nominal conditions are frequently 25 Cͦ and 1 C (one hour discharge rate) [82]. In
the data used in this work capacity measurements were performed at close to these conditions.
𝑆𝑜𝐻 =
𝐶
𝐶𝑛𝑜𝑚
(11)
As mentioned earlier, it is important to keep in mind that the cell capacity is not only a function of the
extent of aging but also temperature and current. In other words, if one wants to accurately compare
the capacities of an aged and a fresh battery, for instance to estimate battery SoH, testing conditions
should strictly speaking be identical. In practice, such as in battery online monitoring, this is rarely
achievable and these effects are therefore often ignored for the sake of practicality. As the nominal
capacity of the fresh battery is normally a known quantity, SoH estimation is essentially a matter of
estimating the capacity of the aged battery. The simplest way of determining the current maximum
capacity of the battery is of course to measure the total transferred charge during a full discharge of
the battery from a fully charged state. This is normally not possible during normal battery operation
and can be time consuming and costly. It also reduces availability of the battery to perform its intended
function [5]. All the issues of Coulomb counting also apply. Naturally, the SoH, as defined in (11), is
closely connected to the SoC through the following relation:
35
𝐶 = 𝑆𝑜𝐻 ∗ 𝐶𝑛𝑜𝑚 =
∆𝑄
∆𝑆𝑜𝐶=
∫ (𝜂𝑖)𝑑𝑡2
1
𝑆𝑜𝐶2 − 𝑆𝑜𝐶1 (12)
Another commonly found SoH definition is the relative increase of internal resistance compared to a
that of a new cell. The easiest way of measuring battery internal resistance is by doing a pulsed charge-
or pulsed discharge test, meaning that a short current signal is charged or discharged to/from the
battery, respectively. The instantaneous voltage response is then equal to the internal resistance
multiplied by the applied current per Ohm’s law [120]. This will be the method of choice for
determining internal resistances in this work and will be described in more detail in a later section.
As a result of the different definitions used, discussions of SoH algorithms may become confusing and
difficult. To alleviate this, efforts have been made to try to conciliate the concept of SoH by establishing
a correlation between capacity loss and resistance increase, thus potentially establishing a universal
definition of the concept. Unfortunately, this work has not been particularly successful, even though
battery-specific correlations certainly seem possible [65, 121]. Looking back at section 1.3, it appears
that the concept of capacity loss encompasses more of the common aging processes and therefore
appears to be a more meaningful metric for diagnosing battery health in terms of its safety status. The
focus will thus largely be on estimating capacity degradation in this work. This is also keeping in mind
the intended application, stationary renewable energy storage, where maximum power is generally of
less importance than maximum energy. For say hybrid electric vehicle applications, a definition based
on internal resistance might be more sensible as issues of power and heat are likely to be more
essential [65, 82].
Determining the SoH is critical for a number of reasons. Firstly, it allows for the prevention of critical
battery failure, which is especially important due to the safety concerns of Li-ion batteries. Secondly,
it allows for the optimization of battery usage and maintenance such as to maximize efficiency and
return on investment [51]. Thirdly, it is necessary for accurate SoC estimation. When effects of
operation conditions on aging for a specific battery chemistry can be predicted, battery systems can
be managed in such a way as to maximize lifetime and thus efficiency and return on investment [10].
SoH is closely linked with the concept of remaining useful life (RuL), the time, or alternatively number
of full equivalent battery cycles, until the battery is no longer considered usable for the intended
application. When RuL is zero the end of life (EoL) has been reached. EoL is in turn usually defined as
when a set SoH has been reached. Note that the battery is always still technically functional at EoL.
The threshold SoH at EoL depends on the intended application. For vehicle applications (EVs, HEVs) it
is often set at 80 percent SoH, while it can be as low as 65 percent for stationary applications [64]. This
means that one can determine the RuL if can predict how SoH will develop over time depending on
certain key contributing aging factors. Alternatively, one can get an idea of the RuL by simple curve
fitting and extrapolation of SoH data over time, as seen in the figure below. This method assumes
identical future operation conditions, of course.
36
Figure 19: Determining RuL by extrapolating SoH measurements [65].
More advanced statistical RuL methods have also been suggested, for instance utilizing a particle filter
[122, 123].
Of course, like SoC, it is not possible to directly measure SoH during battery operation, no matter the applied definition. For this reason, many online SoH estimation methodologies have been, and continue to be, suggested, varying immensely in approach and complexity. Roughly speaking, the most popular online SoH estimation methods are based either on a battery model combined with some filtering- or observer algorithm, or are based on machine learning algorithms, even though combinations have also been suggested [124].
The most common machine learning algorithms are currently based on support vector machines, fuzzy logic, or artificial neural networks. The idea is to develop a “black box” model of the battery using large amount of operational data. The “black box” model can then be used as a virtual copy of the real battery for testing, which can involve internal resistance, capacity, or both. While these types of approaches have the potential of universality and robustness there still exist some issues. For instance, they tend to be too computationally straining in order for them to be implemented in an online controller and require large amount of data to be collected. Instead, operational data is often collected over a time span and then processed in a batch-wise fashion when using these methods [7]. One often speaks of the algorithm “learning” the battery behavior. In one recent work, Klass et al. used approximately 880 000 data points in their SVM-based algorithm, which amounted to an offline training time of 1.6*106 core seconds ≈ 44 hours on the used computer. The data collection has to be repeated whenever a new SoH reading is required [20]. Online, adaptive training therefore seems unfeasible at this stage, even though huge efforts are being put forth in regards to streamlining the computational processes of machine learning methods [7].
In the other category of filtering- and observer based algorithms plenty of suggestions for online solutions have been suggested. These are in generally based on Kalman filtering or sliding mode observers, where Kalman based approaches appear to dominate currently [62]. In general, the struggle is to develop a combination of a battery model, which is always needed for these approaches unlike for machine learning methods, and a regression algorithm that is both sufficiently accurate and computationally efficient. This is a difficult problem in many aspects. Models and algorithms cannot be too detailed as this leads to data points having to be far apart in time which might actually reduce overall estimation accuracy. This means that the battery model and algorithm must be both simple and accurate concurrently.
Plett described methods of SoH determination using both extended Kalman filters and a dual extended Kalman filter but provided little in terms of dynamic validation [125]. The effect of filter tuning on overall performance or estimation accuracy was also not discussed. A simple empirical battery model
37
was used. Kim suggested a combination of a 1 RC ECM and a sliding mode observer algorithm for SoH estimation, thus potentially providing computational advantages over approaches based on Kalman filtering as heavy matrix operation are avoided in this approach [126]. Capacity fade and increase of internal resistance over an accelerated aging test are estimated with good accuracy. Remmlinger et al. used a central difference Kalman filter, incorporating the SoH directly into the state vector. Data from real operation of a hybrid vehicle was used for validation purposes. A definition of SoH based on internal resistance was used and, thus, changes to cell capacity were not considered [127]. Zou et al. developed a method based on a 1 RC ECM and two EKFs, where the two filters operate on different time scales to reduce computational strain compared to the traditional DEKF. The downside to this approach appears to be the need for batch-wise handling of data in the SoH filter. Both capacity fade and increase of internal resistance was estimated [128]. Hu et al. used a very similar approach utilizing time scale separation but with the two filters running online, however only estimating capacity and not internal resistance. Computational strain was compared favorably to the full DEKF. Filter tuning is discussed only briefly [129]. In an extensive recent review, Campestrini et al. compared a large number of different Kalman filtering methods. The focus is on SoC estimation, but the DEKF for online estimation of ECM parameters is also discussed. The issue of filter tuning is brought up as a central issue for all these methods and it is noted that filter tuning is essential in order to achieve as accurate estimation results as possible. Furthermore, it is seen that the optimal filter tuning configuration depends on not only the data used but also the algorithm itself [130].
Due to the seemingly great importance of filter tuning for the performance Kalman filtering based approaches, the need for a thorough analysis of the implications for SoH estimation of Li-ion batteries is evident.
38
2 SCOPE AND AIM OF THESIS
As renewable energy production, such as solar- and wind power, increases in importance and spread,
solutions for reliable, fast and accurate battery monitoring becomes increasingly sought. The currently
dominating battery technology within this field is Li-ion.
This work will examine, compare and contrast Kalman filter based approaches to Li-ion battery
monitoring on the cell level with the focus on SoH estimation. The comparison will be based on the
accuracy, computational performance, robustness and usability of the algorithms. SoC estimation will
be discussed as well due to its near connection to SoH, especially when using Kalman filter based
approaches. Two different SoC estimation algorithms will be analyzed, the extended Kalman filter (EKF)
and the more involved adaptive extended Kalman filter (AEKF) which adaptively updates process- and
measurement noise covariance matrices. SoH estimation, defined as the relative cell capacity
compared to at the start of testing, is performed using three different Kalman filters, comparing joint-
and two different kinds of dual estimation methodologies. The first will be the dual extended Kalman
filter (DEKF), which utilizes two full, integrated EKFs. The two other suggested methods, the single
weight dual extended Kalman filter (SWDEKF) and the enhanced state vector extended Kalman filter
(ESVEKF) are attempts at reducing the complexity of the full DEKF to improve computational
performance. The SWDEKF is a type of dual estimation algorithm where the weight filter state vector
has been reduced in size as to only contain the cell capacity. The ESVEKF is instead a joint estimation
algorithm which utilizes just a single EKF, but with the state vector extended as to contain the cell
capacity. In addition, the effect of ignoring changes to cell internal resistance due to aging while
estimating capacity degradation will be analyzed. Filter tuning, a commonly ignored aspect of Kalman
filtering work, will be given special attention with sensitivity analyses of internal Kalman filter
parameters.
To compare the Kalman filters a simple 1 RC ECM is developed and parameterized. Real battery data,
obtained from the NASA Prognostics Center of Excellence, is used for all model development and
validation of results [9]. The Kalman filters are compared using the same randomized current data set
to simulate realistic operation conditions more accurately than the commonly found in literature
accelerated aging, constant current-constant voltage (CC-CV) full charge-discharge data. All modelling
and simulations will be performed using the MATLAB software.
39
3 METHODOLOGY
Developing an SoH monitoring algorithm based on a Kalman filter method involves several steps. First
a battery model must be developed and validated. In this work a 1 RC ECM was used which means that
the initial parameters of this model: OCV-SoC relationship, resistances, capacitance and cell capacity
must be determined from battery diagnostic data. This involves extraction of these parameters from
current and voltage measurements.
After the battery model has been parametrized, a sufficiently accurate SoC estimator must be
developed. Here two different EKF-based solutions are investigated. Finally, the battery model and SoC
estimator are integrated into the SoH estimator algorithm. Here three different approaches to SoH
estimation are investigated, all based on Kalman filtering. First a dual extended Kalman filter (DEKF)
solution, representing a somewhat standard solution to the problem of SoH estimation, is introduced
[125, 131-133]. Then two potentially less computationally straining SoH estimation algorithms are
suggested. One is based on a single extended Kalman filter with an extended state vector, referred to
as the enhanced state vector extended Kalman filter (ESVEKF) and the other on a dual extended
Kalman filter but with a shortened weight vector consisting of just a single element (cell capacity),
referred to as the single weight dual extended Kalman filter (SWDEKF). All programming and plotting
work was performed in the MATLAB (R2014a and R2017a) software by Mathworks. Example code can
be found in the appendices of the thesis.
The following sections will outline and describe the used methodology in an approximately
chronological order, starting with battery model development.
3.1 EQUIVALENT CIRCUIT BATTERY MODEL DEVELOPMENT To establish an accurate battery model, the behavior of the real battery needs to be characterized
through laboratory testing. The type and number of tests that are needed depends on the complexity
of the intended model. In the case of the battery model used in this work, a 1 RC ECM, the parameters
that need to be specified are the cell capacity, internal resistances R0 and R1, capacitance C1 and the
OCV-SoC relationship. Because the intention in this work is to also estimate the extent of battery aging,
the development of values of parameters relevant to aging over time is also needed as reference
values. To fully parameterize the 1 RC ECM, data from the following types of experiments were used:
Low current (in this case 0.04 A) discharge tests for determination of the cell OCV-SoC
relationship.
Reference discharge tests for periodic determination of cell capacity. This type of test involves
first fully charging the battery using a CC-CV method and then discharging to the cut-off
voltage using a constant current.
Pulsed current discharge tests for periodic estimation of battery transient dynamics, meaning
capacitance and internal resistance parameters.
For all model development and monitoring simulations real battery data is needed. In this work the
“Randomized Battery Usage Data Set” from the NASA Ames Prognostics Data Repository, which is
available for download directly from the NASA Prognostics Center of Excellence website
(https://ti.arc.nasa.gov/tech/dash/pcoe/prognostic-data-repository/) [9], was used for both these
purposes, The purpose of this data set is to approach realistic battery operational data by randomizing
the discharging and charging current in each cycle. In addition, battery diagnostic tests, reference- and
pulsed discharge tests as described above in particular, are performed regularly to allow for tracking
40
of changes in cell behavior over time due to aging processes. Cell voltage, current and temperature
was measured for all tests. Identical tests were performed for several 18650-size Li-ion batteries out
of which the data from the cell RW9 was used unless otherwise stated.
3.1.1 Capacity determination
The cell capacity was measured by discharging a fully charged battery at 2 A constant current until the
lower cut-off voltage of 3.2 V was reached. Before testing for capacity, the battery was charged using
a CC-CV mode, which in this case meant that the battery was charged using a constant current of 2 A
until the upper cut-off voltage of 4.2 V was reached, at which point the voltage was kept constant.
Charging then continued until the current had dropped to 0.01 A. This mode of charging ensures that
the battery becomes very close to fully charged. Determining cell capacity from the constant current
discharge data is then merely a case of a summation of the applied current over all time-steps, I*∆t, of
the complete discharge, in other words by applying Coulomb counting. Coulomb counting could be
used as the initial SoC was known to be at, or very close to, 100 % at the start of every reference
discharge and as measurements were very carefully made using high-precision laboratory equipment,
meaning minimal drift due to sensor bias or other measurement error.
Capacity tests were performed regularly between periods of cycling. The cycling was performed in such
a way as to simulate a random usage pattern, meaning that the applied current was chosen randomly
from a pre-specified set and then applied for five minutes. This procedure was repeated 1500 times
between every capacity measurement. These periods of 1500 cycling steps shall be referred to as
random walk (RW) phases. The exact structure of the RW phases will be explained in more detail in
section 3.2 below. Reference discharge test cell voltage data recorded for battery RW9 over time is
plotted below. To improve the accuracy of the Coulomb counting, a trapezoidal summation method
was used to calculate the cell capacity:
𝐶 = ∑𝐼𝑘 + 𝐼𝑘+1
2∙ ∆𝑡
𝑘
(13)
The sampling time, ∆𝑡 in (13), was ten seconds for all reference discharge tests.
Figure 20: Reference discharge voltage profiles for determination of cell capacity changes over the course of battery testing for RW9.
41
Note in the figure above how upper and lower cut-off voltages did not change during testing. The first
reference discharge test, performed before the start of the first RW phase, is furthest to the right in
the figure above. As the cell capacity fades over time due to various aging phenomena, discharging the
cell from fully a charged state at the same current and cut-off voltages will become faster. As is evident,
cell capacity dropped quite severely during testing. For reference, as mentioned above, for automotive
applications a Li-ion battery is usually considered to be at its EoL when its capacity has faded to 80
percent of its initial nominal value, while 65 % retention of initial capacity is considered acceptable for
stationary applications [134]. The measured cell capacities over the 14 first RW phases of battery RW9
are summarized in table 7 below and are also plotted in figure 21. The capacities given in table 7 are
the average values of the results of the two reference discharge tests between each RW phase. For the
calculation of SoH in table 7, the definition in (11) was used, with the initial capacity before the first
RW phase taken as the nominal capacity. The apparent recovery of cell capacity seen in the
measurement before RW phase 13 is notable. The total loss of capacity over these first 14 RW phases
is more than 35 percent and is thus considered within reason for stationary applications.
Table 7: Measured cell capacities for RW9.
Before RW phase Capacity (Ah) SoH (%)
1 2.10 100
2 2.04 97.1
3 1.98 94.3
4 1.91 90.9
5 1.85 88.1
6 1.76 83.8
7 1.71 81.4
8 1.65 78.6
9 1.60 76.2
10 1.52 72.4
11 1.46 69.5
12 1.42 67.6
13 1.46 69.5
14 1.36 64.8
42
Figure 21: Capacity degradation during testing of RW9. Note the total timespan of around half a year.
3.1.2 Open circuit voltage – state of charge expression
The data used for determination of the OCV-SoC relationship was taken from the low current (0.04 A)
discharge tests. Data from one of these tests is plotted in figure 22 below. Before these tests began
the battery was fully charged in a CC-CV mode, in the same way as described above.
Figure 22: OCV-time relationship for low current discharge test of RW9.
43
The reason that these tests are performed at such a low current is evident when one considers the
suggested 1 RC ECM, representing a reasonable, if simplified, approximation of real battery dynamics:
𝑉𝑡,𝑘 = 𝑂𝐶𝑉𝑘(𝑆𝑂𝐶) − 𝐼𝑘 ∙ 𝑅0 − 𝑉1,𝑘 (14)
𝑉1,𝑘 = 𝑉1,𝑘−1 ∙ 𝑒
−∆𝑡𝑅1∙𝐶1 + 𝑅1 ∙ (1 − 𝑒
−∆𝑡𝑅1∙𝐶1) ∙ 𝐼𝑘−1
(15)
From the equations above, one can draw the conclusion that the expression for cell voltage (14) will
reduce to simply the OCV in the limit of zero current after a certain time. This is because the e-∆t/R1*C1-
term will always be positive and smaller than one, meaning that the transient voltage V1 will decrease
over time and eventually reach zero when no current is applied. As the full OCV-SoC relationship, from
the fully charged state to the lower cut-off voltage, is desired, a low current is applied so that the OCV
is changing, if slowly, while the other terms in (14) remain negligibly small. This allows the OCV to be
measured.
One can then transform the OCV-time plot above to a OCV-SoC form by noticing that the applied
current is of constant magnitude. As transferred charge is simply current multiplied by time one can
multiply the time vector by 0.04 A, the applied current during these tests, to see how much charge has
been emptied from the battery. Finally, 100 % SoC is defined to be the charge in the battery at the
start of this test, representing a fully charged battery, and 0 % SoC is defined to be the point at which
the cut-off voltage, 3.2 V, is reached. The resulting OCV-SoC plot is shown below.
Figure 23: OCV-SoC for RW9 battery.
To utilize this data efficiently in the SoC- and SoH estimation algorithms a least-squares fitting
procedure is suggested, meaning that some function on the form OCV=f(SoC) is fitted to the raw data
such that the sum of the squares of residuals (error) between the fitted curve and the real data points
is minimized. This is easily achieved in the MATLAB environment by using the backslash command. The
only remaining task is then to determine an appropriate form of the function to be fitted to the data.
44
One classic approach is the so-called Nernst equation which on the following form:
𝑂𝐶𝑉 = 𝐾0 + 𝐾1 ∙ ln(𝑆𝑂𝐶) + 𝐾2 ∙ ln (1 − 𝑆𝑂𝐶) (16)
Plett suggested a slightly extended equation, coined the combined model [116]:
𝑂𝐶𝑉 = 𝐾0 + 𝐾1 ∙ ln(𝑆𝑂𝐶) + 𝐾2 ∙ ln(1 − 𝑆𝑂𝐶) −
𝐾3
𝑆𝑂𝐶− 𝐾4 ∙ 𝑆𝑂𝐶 (17)
The issue for both of these very similar equations lies in the nature of battery operation. Batteries will
likely be fully charged and fully discharged several times during their lifetime. In other words, they will
reach the points which should logically be equivalent to 100 % and 0 % SoC, respectively. As one
approaches these points, the logarithmic terms in the equations above will go towards infinity very
rapidly, thus causing the battery model equations to become very unstable and diverge. This problem
is of course made worse by the inevitable existence of measurement- and modeling errors. In the
worst-case scenario, the SoC might briefly jump to a negative value, causing the resulting OCV to be a
complex number when applying (16) or (17). The fact that the cell capacity is continually decreasing
due to aging processes makes these types of phenomena likelier over time as well. The derivatives of
these functions with regards to SoC, necessary for use in extended Kalman filtering, will also display
this behavior. The fully discharged state represented by a SoC equal to zero also presents an issue for
the 1/SoC-term in (17). These equations can therefore be said to display a poor numerical stability and
robustness to small measurement errors.
The suggested solution to these issues is to instead use a high-order polynomial for fitting the data.
This is also commonly done in the literature [111, 135]. In this way one can avoid the issue of instability
due briefly negative SoCs and/or complex values. The cost of this approach is slightly slower
computational performance due to the larger number of coefficients compared to (16) and (17) but
this is considered a small price to pay for the added stability. A 13th degree polynomial was found to
fit the data quite well and was therefore chosen. The need for such a high degree polynomial is evident
when considering the behavior of the fitted polynomials at very low and very high SoC/OCV, where a
good fit becomes critical. The 13th degree polynomial has a maximum error of 0.0292 V which occurs
at zero SoC. The fitted 13th degree polynomial as well as a fitted equation on the form of the combined
model are plotted in figure 24 below. The behavior of the polynomial OCV-SoC outside of 0<SoC<1 is
plotted in figure 24. Note how no spiking issues arise for the polynomial for SoCs over one or under
zero while the combined model behaves strangely at SoCs equal to zero and one.
45
Figure 24: Comparison of polynomial and combined model for OCV-SoC curve fitting.
As can be seen, the fitted equation on the form of the combined model provides a much worse fit. The
comparison is slightly unfair considering the much lower number of fitted coefficients in the combined
model.
Figure 25: Fitted OCV-SoC polynomial is clearly stable while the combined model “spikes”.
It is worth noting that the raw data can also be used directly for SoC estimation by utilizing a look-up
table with linear interpolation between measured data points. This approach was however found to
be much slower than using a fitted expression with little noticeable improvement in terms of accuracy
compared to the polynomial approach and was therefore discarded. Using a fitted function also means
46
that numerical derivation can be avoided when using the model in Kalman filtering, thus improving
performance [136].
3.1.3 Identification of equivalent circuit model parameters
The next step of model development involved identification of the values the ECM parameters R0, R1
and C1. Data from pulsed discharge tests were used for this purpose. For these tests discharged pulses
of 1 A were applied for ten minutes, between each of which the battery was rested for 20 minutes,
during which time no current was applied. At the start of the tests the battery was fully charged in a
CC-CV mode and the procedure was continued until the lower cut-off voltage was reached. The
resulting voltage profiles from these tests are plotted below.
Figure 26: All pulsed discharge tests for determination of ECM parameter values for RW9.
The increase in internal resistance over time can be seen by noticing how the voltage curves in figure
26 are dropping further and further down during the periods of resting. This instantaneous vertical
drop upon the applied current being cut is directly proportional to the internal resistance of the cell,
as seen in figure 27 below.
As all tests were performed at room temperature it was not possible to identify the temperature
dependence of the ECM parameters and therefore the parameter values were assumed to be
independent of the temperature. It is however worth noting, as mentioned previously, that resistance
generally decreases at higher temperatures while the capacity generally increases [113]. This is
especially so when one considers that the cell temperature generally was lower during the diagnostic
steps for capacity and internal resistance estimation than during actual cycling. This was likely due to
the extensive resting of the battery during the diagnostic tests, meaning that Joule heating
(proportional to I2R, see (3)) could largely be dissipated. Notice for instance in figure 31 below how the
temperature rises quite quickly in the beginning of the RW phase and then remains above room
temperature for the duration of RW cycling. This likely also affects the results for later RW phases as
the average temperature during cycling increased over time due to the increased internal resistance
of the cell caused by aging effects.
47
One of the great advantages of the 1 RC ECM is that model resistances and capacitance can be roughly
estimated directly from pulsed discharge data using simple rules of thumb. This is shown in figure 27
below.
Figure 27: Ballpark estimation of ECM parameters from pulsed discharge.
Furthermore, by noting these rules of thumb it is possible to get a basic understanding of how changing
the values of the ECM parameter values affects the behavior of model. While the method of visual
inspection shown in figure 27 certainly is convenient, it naturally lacks accuracy. This is especially so
for identification of the capacitance value as it may prove difficult to determine the exact time at which
Δν∞ has been reached.
It is clear that this is again a question of curve-fitting, just as for determination of the OCV-SoC
relationship. The difference is that here it is not possible to simply use a least squares methodology
directly as the 1 RC ECM is not linear in its parameters. Consider the recursive nature of the transient
voltage V1 for instance. The solution here was to write a MATLAB function incorporating the full 1 RC
ECM and the previously determined OCV-SoC relationship with the resistances and capacitance of the
model as input parameters. The output was then the squared difference between the measured cell
voltage and the cell voltage estimated by the model. The minimum of this function was then found
using the built-in MATLAB function “fminsearch” which uses the Nelder-Mead simplex optimization
algorithm [137]. The resulting fit for the initial pulsed discharge test of RW9 is plotted below.
48
Figure 28: 1 RC ECM model fitting results for first pulsed discharge test of RW9.
As pulsed discharge tests were performed between every other RW phase (every 3000 RW steps), ECM
parameter values could be updated and used as reference for SoH estimation at a later stage. As all
elements of the battery models have now been identified, the result seen in figure 28 above is
representative of the overall model accuracy. The model is evidently not an exact description of the
battery behavior. The modelling error is comparable with reported values for common Li-ion battery
models in the literature [138]. Several authors have extended their ECM by allowing for its parameters
to be SoC-dependent as well [135]. This was not attempted in this work but would likely have led to a
closer model fit. One issue with that approach for this work is that changes to the ECM parameter
values will later be used for SoH estimation. If the ECM parameters themselves are also SoC-dependent
it becomes very difficult to know how to rationally update these appropriately. Also, as will be seen
later, the 1 RC ECM is certainly able to achieve accurate results when used with a Kalman filter
approach despite small modelling errors. In fact, modelling errors are taken into consideration by the
Kalman filter and are thus expected to exist to some extent. The errors in modelling, here defined as
estimated- subtracted by measured cell voltage, of the initial pulsed discharge test of RW9 can be seen
in the figure below.
49
Figure 29: Modeling error in first pulsed discharge test of RW9 using 1 RC ECM.
The modelling error is seen to mostly lie between -30 mV and 20 mV for this case with a root-mean-
square error of 11 mV. It is also notable that the largest absolute modelling errors occur at lower SoCs
towards the left side of figures 29 and 28. Repeating said method of ECM parameter value
identification for data sets taken from later pulsed discharge tests yielded the following development
of the ECM parameters over time for battery RW9:
Table 8: Development of ECM parameters with aging for battery RW9.
Before RW phase R0 [Ω] R1 [Ω] C1 [F]
1 0.074 0.045 815.1
3 0.080 0.066 845.4
5 0.096 0.076 978.8
7 0.113 0.085 1205.4
9 0.133 0.089 1828.4
11 0.149 0.098 1948.6
13 0.169 0.110 2348.7
Both internal resistances parameters R0 and R1 evidently increase by a significant amount during the
course of testing. The capacitance C1 as well. Despite this changes to the capacitance was not
monitored by any Kalman filtering algorithm investigated in this work. This appears to be a common
choice in the literature as well [122, 128]. The reasons for this are two-fold. Firstly, as will be seen
below, changing the capacitance does not reduce modelling error enough to warrant the extra
computational strain required for running the extended monitoring algorithms, and, secondly, no
commonly used definition of battery SoH incorporates the capacitance. Not adapting the capacitance
does introduce some extra modelling error for later RW phases. This is shown in figure 30 below where
the added modeling error for the pulsed discharge test before RW phase 13 is shown. Not adjusting
for capacitance meant using the first measured capacitance value, 815.1 F, for all tests instead of
50
adjusting to the closest capacitance value as estimated via fminsearch, which for the case of RW phase
13 was 2348.7 F for battery RW9 per table 8.
Figure 30: Difference in modelling error when not adjusting for changing capacitance.
The maximum added error by not adjusting the capacitance is 40 mV in figure 30. As can be seen, the
maximum overall error is however the same, approximately 80 mV at around 3.5 hours in to the test.
The overall increase of modelling error for later RW phases might be due to several reasons. The initial
OCV-SoC relation might have been altered, there might be additional temperature and/or C-rate
effects, or phenomena not encapsulated by the simple 1 RC ECM might play a role.
The developed model may be used for estimation of battery state, meaning SoC and transient voltage
V1, and ECM parameter values R0, R1, C1, C, thus allowing for estimation of SoH.
3.2 STATE OF HEALTH ESTIMATION VALIDATION AND CONSIDERATIONS As mentioned previously, no online reference values of capacity and resistance can be measured
during operation of a Li-ion battery. For the data used in this study, tests for determination of internal
resistance and cell capacity were performed periodically, between periods of operation, the so-called
RW phases. In other words, the exact development or path of the cell capacity and internal resistance
over time is not available. Fortunately, these parameter values change only quite slowly, especially in
comparison to state variables like SoC and transient voltage, meaning that these periodic
measurements are reasonable to use as practical references. For methods that attempt to estimate
the cell capacity and/or resistance online, there are two different ways of validating their performance
using this type of data:
1. Let the Kalman filter algorithm run through an entire RW phase with correct, as measured,
starting values of capacity and internal resistances. Compare estimated values of parameters
at end of RW phase simulation with measured values at that point.
2. Set starting guesses of capacity and/or internal resistance substantially off from measured
values intentionally and let the filter estimation converge to the measured values at the start
51
of RW phase. Compare to estimated values of parameters to measured values before RW
phase.
The first of these approaches has the advantage of showing the full estimated development of capacity
and/or internal resistance over the course of battery operation. This would be representative of a long-
term battery monitoring situation. The computation time needed for such a simulation might however
be quite extensive, depending on what hardware and software is used, the length of operation
between reference measurements, and the data sampling time. Many measurements taken over a
long period of time need to be considered for validation of SoH-estimation methods in order for
changes to the ECM parameters to be clearly distinguishable. The second method solves this issue of
time consumption as data from only the first part of a given RW phase is needed for the simulation.
This is also representative of the common case of ECM parameters not being known exactly and
needing to be estimated online. In addition, the existing SoC reference, Coulomb counting, uses the
capacity measured before a given RW phase in all calculations, meaning that the quality of this
reference is worsened towards the end of a given RW phase when cell capacity will have decreased
due to aging. Both validation methods will be used in this study, but the second more so due to the
quicker results. The exact nature of the RW operation during cycling of the batteries will now be
outlined.
Between measurements of cell capacity and transient behavior, referred to as reference- and pulsed
discharge tests respectively, the battery is operated in a so-called random walk (RW) cycling mode.
The structure of the data set is as follows: first reference-, pulsed-, and low current discharge tests are
performed to determine initial battery capacity, OCV-SoC relation and transient dynamics. Then 1500
RW steps are performed per the following template:
1. Fully charge the battery in a CC-CV mode.
2. Pick a discharging/charging current randomly from the set [-4.5 A, -3.75 A, -3 A, -2.25 A, -1.5
A, -0.75 A, 0.75 A, 1.5 A, 2.25 A, 3 A, 3.75 A, 4.5 A]. Here, as throughout this work, negative
currents indicate charging and positive currents discharging.
3. The randomly picked current is applied for 5 minutes or until the battery voltage is on the
edge of the battery SOA, meaning 3.2 V or 4.2 V. This will be referred to as one RW step from
here on out, regardless of duration.
4. After each RW step there is a short rest for under 1 second during which no current is applied
to the battery.
a. If 1500 RW steps, referred to as one full RW phase for here on out, have been
performed, end RW cycling and proceed to diagnostic testing.
b. Otherwise go to step 2 again.
This type of cycling data will be used for all simulations of SoC- and SoH monitoring. Typical
measurements of cell voltage, current and temperature for the first 100 steps of a RW phase are
displayed in figure 31. This particular example comes from the first RW phase of battery RW9.
52
Figure 31: Voltage, current and temperature over the first 100 RW steps for RW9.
The average duration of a RW step was quite a bit less than five minutes, as one full RW phase took
around 95 hours on average. The experiments were performed at room temperature. The elevated
temperature during RW cycling, likely due to Joule heating, is evident. This mode of operation is very
intense as batteries are virtually constantly being either charged or discharged. This means that aging
effects will be accelerated in these tests compared to most normal Li-ion battery applications.
It was discovered that negative SoC would quickly result if Coulomb counting was applied to data from
later RW phases of RW9, even if the cell capacity used in the calculations was adjusted to the latest
value as measured in the reference discharge tests. See for instance the SoC calculated by Coulomb
counting for RW phase 10 of battery RW9 in the figure below. The calculated SoC is as low as -10 % at
just after eight hours into the RW phase.
Figure 32: Negative SoC by Coulomb counting in RW phase 10 for battery RW9.
53
The reason for this has not been confirmed but might have to do with the fact that all reference
discharge tests for capacity determination were performed at 2 A constant current and at generally
lower temperatures than during RW cycling. It is widely known that higher temperatures and lower
currents correlate with a higher cell capacity, meaning that the real capacity might be higher during
RW phases than during reference discharge tests [82].
Having outlined the procedure of model development and special considerations regarding SoH
estimation, the utilized Kalman filtering regression algorithms used for SoC and SoH estimation will
now be explained in detail.
3.3 KALMAN FILTERING Kalman filtering, as briefly outlined in the introduction of this work, is a family of recursive algorithms
for online or batch estimation of internal, non-measurable, variables of a system with system inputs-
and outputs having a certain noise or uncertainty associated with them. The goal is to minimize the
mean-square estimation error of these internal variables for a stochastic system. For this reason, the
Kalman filter may be seen as the stochastic, recursive equivalent of the deterministic, batch least
squares method [139]. The Kalman filter has many applications such as in process control, navigation,
and tracking [140]. One major advantage of the Kalman filter methodology is that the error bounds of
the internal variable estimates are always calculated concurrently. One frequently mentioned
downside is the large number of matrix operations needed, leading to quite low computational
performance, especially as the number of estimated state variables increases [116, 141]. In fact, the
application of Kalman filtering methods has largely been driven by advances in available computing
power [142]. Overcoming, or at least handling of, this issue within the framework of Li-ion battery
monitoring will be central in this work.
The basic principle of Kalman filtering is to construct a model of the real system which uses the internal
variables which one wishes to estimate as inputs, while the outputs are some measurable variables.
The output (may be a vector of values) of the model for a given system state and input is compared to
the measurement. The internal variables are then updated according to the difference between the
estimate and the measured value: if the difference is large than the update is large and vice versa. At
the same time, measurement- and process noise is incorporated directly into the algorithm. If the
measurement noise is large, updates will be smaller relative to the error and vice versa. If the
measurement noise is large, the measurements are considered less important, or “less trusted”, than
the model prediction. The process noise represents unmodeled inputs to the system, such as random
perturbations, and modeling errors. The calculated Kalman gain matrix controls the balance of
weighing between the measured values and those predicted by the model. In the following
introduction to the underlying principles of Kalman filtering all equations will refer to the original
algorithm developed by Rudolf Kalman during the 1950s, frequently referred to as the linear Kalman
filter (LKF), for ease of explanation and because of the similar fundamental algorithm structures pf all
Kalman filtering approaches [143]. More advanced Kalman varieties will be introduced at a later stage.
The basic Kalman filter principle may be represented graphically as so:
54
Figure 33: The Kalman filtering principle [116].
In figure 33 uk represents some deterministic system input. The real, non-measurable system state is
represented by the internal variables xk, which is to be estimated by the algorithm. The deterministic
system input is also sent to the system model, which results in the estimated internal variables 𝑥𝑘 and
the estimated system output �̂�𝑘. The estimated output is then compared to the measured output and
the difference between the measured and estimated output, often called the error, is used to update
the estimate of the system state, weighted by the calculated Kalman gain matrix Lk, which also
considers current state uncertainty and measurement noise. Details of the algorithm will now be
explained step-by-step.
The core of the Kalman filter methodology consists of two equations, referred to here as the state
equation (18) and the output equation (19). For the LKF these can be written as:
𝑥𝑘+1 = 𝐴𝑘 ∙ 𝑥𝑘 + 𝐵𝑘 ∙ 𝑢𝑘 + 𝑤𝑘 (18)
𝑦𝑘 = 𝐶𝑘 ∙ 𝑥𝑘 + 𝐷𝑘 ∙ 𝑢𝑘 + 𝑣𝑘 (19)
Here k is the time index, x the state vector, u the deterministic system input, and y the measurable
output. As an example, x will be a vector of two elements in this work, consisting of the state variables
SoC and transient voltage. The measurement, y, will be the cell voltage. The current will represent the
deterministic system input u. The matrices A, B, C and D describe the dynamics of the system and
contain the model coefficients and structure. These matrices may also be time dependent. Vectors w
and v represent process- and measurement noise of the system respectively. The equations (18) and
(19) may together be referred to as a state-space representation of the system. It is notable that w and
v must be specified a priori as part of filter initialization, often a difficult task. Furthermore, w and v
are assumed to follow a certain set of assumptions: they are zero-mean and Gaussian (normally
distributed) and also mutually uncorrelated [135, 142]. While these assumptions are likely rarely
completely fulfilled, the results of the Kalman filter have spoken for themselves [116]. In any case, it is
very useful to be able to model both measurement- and process noise during filter tuning, an activity
which will be discussed in depth at a later stage in this study. The process- and measurement noises
have covariance matrices Σw and Σv respectively, which are used directly in the filter algorithm.
The Kalman filter algorithm can essentially be divided into three stages: initialization, the time update
the measurement update. First the filter must be initialized before monitoring can commence. The
initialization stage requires the user to specify initial guesses of the system state, 𝑥0+ , along with the
associated initial state covariance matrix, Σx,0 and the previously mentioned process- and
measurement noise covariance matrices Σw and Σv. Note that from here on out estimated values will
be specified with the hat symbol while measured values will not have the hat to differentiate these
55
from each other, just like in figure 33 above. The initial state estimate and state covariance matrix are
quite straightforward to specify, Σx,0 usually being set as the following diagonal matrix2 [144]:
Σ𝑥,0 = 𝑑𝑖𝑎𝑔(𝔼[(𝑥0 − 𝑥0+)(𝑥0 − 𝑥0
+)𝑇]) (20)
Here 𝔼 is the statistical expectation value operator and T represents the vector transpose. The initial
guess of the system state is based on the best available information and experience. Of course, the
true value of x0 is still generally not known. This is however not a major issue for most cases for two
reasons. Firstly, one will usually have some rough idea of where x0 might lie. This is certainly the case
for battery modelling as shall be seen later. It is generally best to assume that the initial guess is as
poor as possible, meaning that the guess is at one end of the range of thinkable values while the real
value happens to be at the other end. Secondly, the initial state covariance matrix will only be used for
the first iteration. As new information in the form of measurements come in, the state covariance is
updated recursively to reflect the model uncertainty. The state covariance matrix will not necessarily
remain diagonal during filtering but will always be positive definite and symmetric [143, 144]. As the
model converges towards true estimates of the system state, the diagonal elements of the state
covariance matrix can be seen as a measurement of model or state estimate accuracy, meaning that
error bounds are automatically updated during filtering. As a rule of thumb, the true value of x will be
bounded by the following expression around 95 % of the time if the underlying assumptions of the
Kalman filter are fulfilled [116]:
𝑥 = 𝑥𝑘 ± 2 ∙ √𝑑𝑖𝑎𝑔(Σ𝑥,𝑘) (21)
Unlike the state covariance matrix, the process- and measurement noise covariance matrices will
remain constant during filtering. These two parameters represent the inherent uncertainty, or
inaccuracy, of the model (values of 𝑥) and measurements (values of y) respectively. As these are not
updated recursively in the standard Kalman filtering methods, finding suitable values for these
becomes crucial to achieve acceptable filter performance. The measurement noise matrix values can
be estimated readily if earlier measurement data is available. For the case of a number of
measurement signals i with time-independent standard deviations σi, the measurement noise
covariance matrix can be estimated as a diagonal matrix:
Σ𝑣 = 𝑘 ∙ 𝑑𝑖𝑎𝑔(𝜎𝑖2) (22)
Where k is most often one or slightly larger than one and can be seen as a tuning parameter, say if one
expects larger measurement error in the future k can be made larger. If the measurement error matrix
is too large, the information gathered from measurements will be ignored and the filter will just use
the model predictions which can cause the filter to diverge. Setting the measurement error covariance
matrix too small will cause the filter to fail to separate out sensor noise from true signal, causing
reduced state estimation accuracy [144].
Most often it is the specification of the process noise covariance matrix values which represents the
biggest challenge in filter tuning. Several methods have been suggested to try to rationalize the
process. Because of the relative difficulty, a trial-and-error procedure is often necessary [144]. The
process noise covariance matrix is in any case diagonal and time-independent in standard Kalman filter
approaches. In general, finding the appropriate values of Σ𝑣 and Σ𝑤 will be done in one of three
possible ways [145]:
2 A matrix is called diagonal if all values not on the diagonal are zero.
56
1. Fixing Σ𝑣 , given that a reasonable estimate of it exists, and vary Σ𝑤 by trial and error until
realistic, stable results are yielded.
2. The reverse situation, meaning that Σ𝑤 is known and Σ𝑣 is varied in a trial-and-error fashion.
3. Neither Σ𝑣 nor Σ𝑤 are known and both must be found by trial-and-error by varying both.
In this work the relative size of the values in the process noise covariance matrix was found to be an
important factor in achieving accurate state estimation. This will be discussed further in the results
section.
Finally, the LKF algorithm proper can be introduced. Note that the algorithm equations will not derived
here. If the reader is interested she may see the original article by Robert Kalman for such a derivation
[143]. After the initialization step, the time update can be commenced. Parameters will be notated
with a minus (“-“) sign index before the measurement update and with a plus (“+”) index afterwards
to differentiate these. The time update step essentially propagates the state and state covariance to
the next time step through the model equations. It may thus be considered the model prediction step.
State estimate time update: 𝑥𝑘− = 𝐴𝑘−1 ∙ 𝑥𝑘−1
+ + 𝐵𝑘−1 ∙ 𝑢𝑘−1 (23)
State covariance matrix time update: Σ𝑥,𝑘− = 𝐴𝑘−1 ∙ Σ𝑥,𝑘−1
+ ∙ 𝐴𝑘−1𝑇 + Σ𝑤 (24)
Calculation of Kalman gain matrix:
𝐿𝑘 =Σ𝑥,𝑘
− ∙ 𝐶𝑘𝑇
𝐶𝑘 ∙ Σ𝑥,𝑘− ∙ 𝐶𝑘
𝑇 + Σ𝑣
(25)
Note that the state estimate time update equation assumes zero process noise. In other words, it is
deterministic. If the system is stable, the state covariance matrix should decrease over time. The
process noise covariance matrix will always lead to larger values in the state covariance matrix, thus
increasing state uncertainty. In (25) the Kalman gain matrix is strongly dependent on both the state
covariance- and measurement noise covariance matrices. If the state covariance matrix is large, then
the Kalman gain matrix will be large and the state estimate measurement update will be more drastic
as seen below. At the same time, a large measurement error covariance matrix will lead to a small
Kalman gain matrix and thus the opposite effect. In fact, the following two limits are noted [142]:
limΣ𝑣→0
𝐿𝑘 =1
𝐶𝑘 (26)
limΣ𝑥,𝑘
− →0𝐿𝑘 = 0 (27)
In other words, as the measurement error approaches zero, the measurements are more “trusted”.
Likewise, when the state covariance matrix, representing the state estimate uncertainty, approaches
zero, model predictions are more “trusted”. The Kalman gain matrix may therefore be seen as
something of a signal-to-noise metric or, as mentioned earlier, the relative trust put in measurements
versus model predictions [116].
57
In the measurement update stage, the new measurement, yk, is utilized to update the values of the
state estimate and state covariance matrix. The measurement update step may be considered a
correction step where the predicted system state is adjusted by utilizing the new measurement data.
Output estimate error: 𝑒𝑘 = 𝑦𝑘 − �̂�𝑘 = 𝑦𝑘 − (𝐶𝑘 ∙ 𝑥𝑘− + 𝐷𝑘 ∙ 𝑢𝑘) (28)
State estimate measurement update: 𝑥𝑘+ = 𝑥𝑘
− + 𝐿𝑘 ∙ 𝑒𝑘 (29)
State covariance matrix measurement update: Σ𝑥,𝑘+ = (𝐼 − 𝐿𝑘 ∙ 𝐶𝑘) ∙ Σ𝑥,𝑘
− (30)
If the output estimate error is large it is more likely that estimated internal variables have been
inaccurately estimated and adjustments to these should therefore be large. This is reflected in (29).
Note in (30) how the state covariance matrix always gets smaller due to new measurement
information. Here I in (30) represents the identity matrix.
The time- and measurement update steps are then repeated in every time step for the duration of the
monitoring process. The six steps in the time update and measurement update steps are essentially
the same for all Kalman filter approaches with only minor modifications for some more advanced
varieties. Having introduced the Kalman filter basics through the example of the LKF, a series of
modified, more advanced methods are now presented. The purpose of these more advanced Kalman
filters is either to improve the estimation accuracy for non-linear systems or, in the case of the adaptive
extended Kalman filter, a streamlining of the filter tuning process. The last three Kalman filtering
methods to be introduced are SoH estimators, meaning that also the cell capacity and, in the case of
the DEKF, the internal resistance parameters R0 and R1 are estimated concurrently with the SoC.
3.3.1 Extended Kalman filter
If one studies the state- and output equations ((18) and (19)) one will notice that these describe a
linear system, thus the name LKF. Because of this structure, linear systems are a requirement for using
the LKF. Batteries are however strongly non-linear systems and their modelling should therefore
preferably reflect this. Indeed, if the model suggested in this work, the 1 RC ECM, is examined it is
found that a direct separation of the model equations into matrices A, B, C and D such as in (18) and
(19) is not possible. The simple and widely accepted solution to this problem is to linearize the model
equations via a truncated Taylor series expansion. After this series expansion has been performed, the
same algorithm as in the LKF may be used. This is called extended Kalman filtering (EKF). The state-
and output equations are now assumed to be non-linear:
𝑥𝑘+1 = 𝑓(𝑥𝑘 , 𝑢𝑘) + 𝑤𝑘 (31)
𝑦𝑘 = 𝑔(𝑥𝑘 , 𝑢𝑘) + 𝑣𝑘 (32)
Taylor series expansion allows for the linearization of these functions. All terms of the series but the
first two are truncated:
𝑓(𝑥𝑘 , 𝑢𝑘) ≈ 𝑓(𝑥𝑘 , 𝑢𝑘) +
𝛿𝑓(𝑥𝑘 , 𝑢𝑘)
𝛿𝑥𝑘(𝑥𝑘) ∙ (𝑥𝑘 − 𝑥𝑘) (33)
𝑔(𝑥𝑘 , 𝑢𝑘) ≈ 𝑔(𝑥𝑘, 𝑢𝑘) +𝛿𝑔(𝑥𝑘 , 𝑢𝑘)
𝛿𝑥𝑘(𝑥𝑘) ∙ (𝑥𝑘 − 𝑥𝑘) (34)
Depending on how non-linear the system is this might be a more or less drastic approximation [140].
As will be shown, using the first two terms of the Taylor expansion will prove to be sufficient for the
58
purposes of battery modelling here, though some authors have suggested more sophisticated
solutions to the problem of model non-linearity such the unscented Kalman filter or the particle filter
[146, 147]. Of course, the EKF requires the model equations f and g to be differentiable at all points.
By simple comparison of the LKF and EKF state- and output equations it is evident that what must be
determined for the LKF algorithm to be applicable for this non-linear system are equivalents of the
matrices A, B, C and D. From an analytical standpoint, the A and C matrices describe how the current
system state affects future system states and outputs, respectively. By comparison of (34) and (33)
with (19) and (18), it is evident that the equivalents of these equations must be the Jacobian matrices.
Therefore, these will be renamed accordingly:
�̂�𝑘+1 =
𝛿𝑓(𝑥𝑘 , 𝑢𝑘)
𝛿𝑥 (35)
�̂�𝑘 =
𝛿𝑔(𝑥𝑘 , 𝑢𝑘)
𝛿𝑥 (36)
If (35) and (36), along with (33) and (34), are inserted back into (31) and (32) the following equations
are the result:
𝑥𝑘+1 = �̂�𝑘 ∙ 𝑥𝑘 + 𝑓(𝑥𝑘 , 𝑢𝑘) − �̂�𝑘 ∙ 𝑥𝑘 + 𝑤𝑘 (37)
𝑦𝑘 = �̂�𝑘 ∙ 𝑥𝑘 + 𝑔(𝑥𝑘 , 𝑢𝑘) − �̂�𝑘 ∙ 𝑥𝑘 + 𝑣𝑘 (38)
The choices of A- and C matrix equivalents are therefore validated as they serve the same functions in
(37) and (38) as A and C did in (18) and (19), respectively. The parts of (37) and (38) which are not
dependent on x then functionally replace the terms which are multiplied by the matrices B and D in
(18) and (19). Now the exact same procedure as in the LKF, meaning the six time- and measurement
update equations, may be used by incorporating these modified A and C matrices. This allows for
Kalman filtering to be utilized even though the model equations are non-linear.
It is now easy to write the 1 RC ECM model equations in the matrix form appropriate for the EKF. The
state equation shall be the following, utilizing the notation in figure 18:
𝑥𝑘+1 = [𝑉1,𝑘+1
𝑆𝑂𝐶𝑘+1] = [𝑒
−∆𝑡𝑅1∙𝐶1 00 1
] ∙ [𝑉1,𝑘
𝑆𝑂𝐶𝑘] + [
𝑅1 ∙ (1 − 𝑒−∆𝑡𝑅1∙𝐶1)
∆𝑡
𝐶
] ∙ 𝐼𝑘 (39)
𝑦𝑘 = 𝑉𝑡 = 𝑂𝐶𝑉(𝑆𝑂𝐶𝑘) − 𝐼𝑘 ∙ 𝑅0 − 𝑉1,𝑘 (40)
This means that the state vector will consist of two elements in this case: the SoC and V1, the transient
voltage over the lone RC pair of the ECM, while the output vector will consist of a single element: the
cell terminal voltage. As the output vector is only of length one, the measurement noise covariance
matrix will also consist of a single element. It will therefore be referred to as the measurement noise
covariance from here on out. Similarly, the Kalman gain matrix will simply be a vector of length two.
The deterministic input is the charging or discharging current, also a single element vector. Finally, by
using (36) with the 1 RC ECM output equation in (40), the C matrix of the EKF may be written as:
�̂�𝑘 = [−1𝛿(𝑂𝐶𝑉)
𝛿(𝑆𝑂𝐶)] (41)
59
Of course, as the chosen expression for the OCV-SoC relation is a simple polynomial, as described in
section 3.1.2, the calculation of the partial derivative of the OCV with regards to SoC in (41) is
straightforward and is easily done analytically.
3.3.2 Adaptive extended Kalman filter
One of the major hurdles when developing a Kalman filtering method is the process of filter tuning,
particularly the identification of suitable process- and measurement noise covariance matrices. The
AEKF attempts to overcome this problem by recursively updating these matrices as new
measurements are made available during filtering. Being a variant of the EKF, the structure of the AEKF
is very similar overall. The differing equations come in the measurement update stage of the algorithm.
First the squares of the output estimation errors are summarized over a certain number of the latest
time-steps, called the moving estimation window size, M. Then updated process- and measurement
noise covariance matrices are updated accordingly [130, 145, 148]:
𝐻𝑘 =1
𝑀∑ 𝑒𝑖 ∙ 𝑒𝑖
𝑇
𝑘
𝑖=𝑘−𝑀+1
(42)
Σ𝑣 = 𝐻𝑘 + �̂�𝑘 ∙ Σ𝑥,𝑘
− ∙ �̂�𝑘𝑇 (43)
Σ𝑤 = 𝐿𝑘 ∙ 𝐻𝑘 ∙ 𝐿𝑘
𝑇 (44)
Note that while one now does not need to specify process- and measurement noise covariance
matrices explicitly, starting guesses for these still need to be specified. This actually means that more
variables need to be pre-specified during initialization of the AEKF than of the EKF as starting guesses
for process- and measurement noise covariance matrices are still needed along with the additional
moving estimation window size. The adaptive updating of the process- and measurement noise
covariance matrices should however make the filter more robust with regards to initialization and thus
reduce the risk of the filter diverging. Before M time-steps have been performed the starting guesses
for process- and measurement noise covariance matrices are used instead of (43) and (44).
3.3.3 Dual extended Kalman filter
The EKF and AEKF algorithms as presented thus far are SoC estimators. During filtering, all model
parameters will be assumed to remain constant when using these filters. This is an approximation as
the capacity, capacitance and internal resistance of the battery will change depending on operating
conditions and most importantly aging. While the EKF and AEKF can handle certain model inaccuracy,
the quality of the state estimation will degrade relatively rapidly if changes to parameter values are
ignored. In addition, tracking of the changes to capacity and internal resistance is of interest in and of
itself as these parameters are essential to monitoring of battery performance and safety. To achieve
this the DEKF is introduced. The idea here is to run two EKFs in parallel, with one normal filter, the
state filter, estimating the system state variables and one so-called weight filter that is estimating the
model parameters concurrently. The filters are run in tandem in an intelligent way such that certain
useful interconnections are utilized for increased computational performance [149].
The weight filter will have its own process- and measurement noise covariance matrices, starting
guesses of the parameter values (weights) and weight covariance matrix, separate to those of the state
filter. The output equation will however be the same for both filters. The weight time update equation
will be very simple. The weight filter state equation is modelled to be slowly changing by applying a
certain process noise, r:
60
𝜃𝑘+1 = 𝜃𝑘 + 𝑟𝑘 (45)
The reason for this simple model is the difference in rate of change between the state variables, such
as SoC and transient voltage, and the elements of the weight filter, such as cell capacity and internal
resistance parameters. With regards to the state variables the weights are practically constant.
Moreover, no accepted, universal equations exist for the changes in capacity or internal resistance
over time.
As in the name, both filters in the DEKF are of the extended kind, meaning that Jacobian matrices are
required for the linearization procedure. The Jacobians A and C for the state filter will remain the same
as those in the EKF as the state- and output equations have not changed. The next step is to determine
what these matrices will be for the weight filter. To differentiate the filters and their internal
parameters, the subscripts x and w will be used for the state- and weight filter respectively.
First it is evident that the weight filter state equation, unlike the case for the state filter, is in fact linear
in the parameters, with Aw being simply the identity matrix. Bw is naturally zero as the parameters are
independent of the applied current. Since the output equation of the weight filter is the same as for
the state filter, Cw will be a Jacobian matrix which needs to be calculated from the Taylor series
expansion, just as for Cx above with the difference that the derivative is with regards to 𝜃 instead of x.
What one will find when one starts to try to calculate Cw is that a recursive procedure is necessary. Like
the Cx matrix is the partial derivative of the output function with regards to x, Cw will be the total
derivative of the output equation with regards to 𝜃:
�̂�𝑤 =
𝑑𝑔(𝑥𝑘−, 𝑢𝑘 , 𝜃)
𝑑𝜃(𝜃𝑘)
(46)
Now note that the output function g is a function of 𝑥𝑘− which is indeed a function of θ, but also that
𝑥𝑘− is a function of 𝑥𝑘−1
+ via the time update equation as well, which in turn is also a dependent on θ.
This is why the total derivative and not the partial derivative with regards to θ is needed for the
calculation of Cw3 [149]. The recursive nature of these calculations is evident, meaning that the last
calculated value of Cw is needed for the calculation of the next one. Cw will simply be initialized as zero,
using the following set of recursive equations, derived from the total derivation of the output equation,
for calculating following values:
𝑑𝑔(𝑥𝑘
−, 𝑢𝑘 , 𝜃)
𝑑𝜃=
𝛿𝑔(𝑥𝑘−, 𝑢𝑘 , 𝜃)
𝛿𝜃+
𝛿𝑔(𝑥𝑘−, 𝑢𝑘 , 𝜃)
𝛿𝑥𝑘− ∙
𝑑𝑥𝑘−
𝑑𝜃 (47)
𝑑𝑥𝑘
−
𝑑𝜃=
𝛿𝑓(𝑥𝑘−1+ , 𝑢𝑘−1, 𝜃)
𝛿𝜃+
𝛿𝑓(𝑥𝑘−1+ , 𝑢𝑘−1, 𝜃)
𝛿𝑥𝑘−1+ ∙
𝑑𝑥𝑘−1+
𝑑𝜃 (48)
𝑑𝑥𝑘−1
+
𝑑𝜃=
𝑑𝑥𝑘−1−
𝑑𝜃− 𝐿𝑘−1
𝑥 ∙𝑑𝑔(𝑥𝑘−1
− , 𝑢𝑘−1, 𝜃)
𝑑𝜃 (49)
Note how (48) originates from the state filter time update equation, and (49) from the state filter
measurement update equation (29).
This system of recursive equations hold under the assumption that the Kalman gain, L, us not a function
of the weight vector θ. While this is not strictly true, it is usually not worth the extra computational
3 Note that the reverse is of course not true: θ is not a function of x when using the set of equation in this work. The partial derivative is therefore sufficient for the calculation of Cx.
61
effort to calculate the extra derivatives needed in every iteration considering how small the difference
in the results is [149]. While the set of equations above may appear intimidating, a lot of results in one
iteration may be re-used in the next thus greatly simplifying the overall procedure.
In fact, one will find that only a relatively small number of new derivatives must be calculated in each
time-step. Firstly, note that the already calculated Cx- and Ax matrices from the state filter may be used
directly in (47) and (48) like so by noting the expressions for these in (35) and (36) above:
𝑑𝑔(𝑥𝑘
−, 𝑢𝑘 , 𝜃)
𝑑𝜃=
𝛿𝑔(𝑥𝑘−, 𝑢𝑘 , 𝜃)
𝛿𝜃+ �̂�𝑘
𝑥 ∙𝑑𝑥𝑘
−
𝑑𝜃 (50)
𝑑𝑥𝑘
−
𝑑𝜃=
𝛿𝑓(𝑥𝑘−1+ , 𝑢𝑘−1, 𝜃)
𝛿𝜃+ �̂�𝑘 ∙
𝑑𝑥𝑘−1+
𝑑𝜃 (51)
The Kalman gain matrix in the previous step, 𝐿𝑘−1𝑥 , is of course also already calculated and just needs
to be stored for the next iteration of the algorithm. This re-usage of already calculated results also
applies for the full derivatives Cw and 𝑑𝑥−
𝑑𝜃 as well. The result is that only the Jacobians
𝛿𝑓
𝛿𝜃 and
𝛿𝑔
𝛿𝜃 need
to be calculated in each iteration. In the case of Li-ion battery SoH monitoring with a 1 RC ECM, the
following weight vector will be used in the DEKF:
𝜃𝑘 = [𝑅0 𝑅1 𝐶] (52)
Note again how the capacitance is assumed to remain constant during cell operation and is therefore
not included into the weight vector. The needed Jacobians may then be written thusly, noting that only
the V1-term in the output function depends on R1 and likewise that only the OCV-term depends on the
cell capacity:
𝛿𝑔(𝑥𝑘−, 𝑢𝑘 , 𝜃)
𝛿𝜃= [
𝛿𝑔(𝑥𝑘−, 𝑢𝑘 , 𝜃)
𝛿𝑅0
𝛿𝑔(𝑥𝑘−, 𝑢𝑘 , 𝜃)
𝛿𝑅1
𝛿𝑔(𝑥𝑘−, 𝑢𝑘 , 𝜃)
𝛿𝐶]
= [−𝐼𝑘 0𝛿𝑂𝐶𝑉(𝑥𝑘
−, 𝑢𝑘 , 𝜃)
𝛿𝐶]
(53)
𝛿𝑓(𝑥𝑘−1
+ , 𝑢𝑘−1, 𝜃)
𝛿𝜃=
[ 0
𝛿𝑉1(𝑥𝑘−1+ , 𝑢𝑘−1, 𝜃)
𝛿𝑅10
0 0𝛿𝑆𝑂𝐶(𝑥𝑘−1
+ , 𝑢𝑘−1, 𝜃)
𝛿𝐶 ]
(54)
Note in the Jacobians above that the state equation f is completely independent of R0. By noting the 1
RC ECM model equations, the partial derivative for the transient voltage with regards to R1 can be
written as the following by utilizing the product and chain rules:
𝛿𝑉1(𝑥𝑘−1
+ , 𝑢𝑘−1, 𝜃)
𝛿𝑅1= 𝑉1,𝑘−1 ∙ 𝑒
−∆𝑡𝑅1∙𝐶1 ∙ ∆𝑡 (𝐶1 ∙ 𝑅1
2)⁄ + 𝐼𝑘−1 ∙ ((1 − 𝑒−∆𝑡𝑅1∙𝐶1) +
−∆𝑡 ∙ 𝑒−∆𝑡𝑅1∙𝐶1
𝐶1 ∙ 𝑅1) (55)
The derivative of the SoC with regards to the cell capacity is also needed. It may be written as the
following by using the expression for SoC in (39):
𝛿𝑆𝑂𝐶(𝑥𝑘−1
+ , 𝑢𝑘−1, 𝜃)
𝛿𝐶=
𝐼𝑘−1 ∙ ∆𝑡
𝐶2 (56)
62
If one should want to include the final ECM parameter, the capacitance C1 in the weight vector, the
partial derivative of V1 with regards to C1 would also be required. This is easily derived to be:
𝛿𝑉1(𝑥𝑘−1
+ , 𝑢𝑘−1, 𝜃)
𝛿𝐶1=
(𝑉1,𝑘−1+ − 𝑅1 ∙ 𝐼𝑘−1) ∙ ∆𝑡 ∙ 𝑒
−∆𝑡𝑅1∙𝐶1
𝐶12 ∙ 𝑅1
) (57)
The final partial derivative in (53), that of the OCV with regards to the capacity, may be taken as zero
under the assumption that the OCV-SoC relation is independent of cell aging. Now that all Jacobians
and their containing partial derivative have been determined the overall structure of the DEKF may be
summarized. Again, it is essentially a question of two EKFs running in parallel.
The first step is the initialization of both filters. Process- and measurement noise covariance matrices
for both filters, initial state- and weight filter covariance matrices, and starting guesses for state- and
weight vectors all must be specified. The general interpretations of the effects of these internal
parameters on the behavior of the filters are the same as in the EKF, perhaps except for the weight
filter process- and measurement noise covariance matrices due to the lack of a more rigorous weight
filter state equation. The choice of internal parameter values for the weight filter is especially critical
for this reason [130].
After initialization, the following steps will be performed in order: time update for weight filter, time
update for state filter, measurement update for state filter, and lastly measurement update for weight
filter. After that the cycle is repeated as such until the end of filter operation. The state filter time- and
measurement update steps are identical to those in the EKF and shall therefore not be repeated here.
The weight filter time update step comprises the following equations:
𝜃𝑘
− = 𝜃𝑘−1+ (58)
Σ𝜃,𝑘
− = Σ𝜃,𝑘−1+ + Σ𝑟 (59)
𝐿𝑘𝜃 =
Σ𝜃,𝑘− ∙ (𝐶𝑘
𝜃)𝑇
𝐶𝑘𝜃 ∙ Σ𝜃,𝑘
− ∙ (𝐶𝑘𝜃)𝑇 + Σ𝑒
(60)
Here Σ𝑟 is the weight filter process noise covariance matrix and Σ𝑒 the weight filter measurement
noise covariance matrix. Note in (59) that one might certainly write this on the same form as for the
equivalent equation in the state filter time update (24), that is with the Aw matrix. This is however
omitted here for convenience as Aw is simply the identity matrix. The measurement update step
equations for the weight filter are the following:
𝑒𝑘 = 𝑦𝑘 − �̂�𝑘 = 𝑦𝑘 − (𝐶𝑘 ∙ 𝑥𝑘
− + 𝐷𝑘 ∙ 𝑢𝑘) (61)
𝜃𝑘
+ = 𝜃𝑘− + 𝐿𝑘
𝜃 ∙ 𝑒𝑘 (62)
Σ𝜃,𝑘
+ = (𝐼 − 𝐿𝑘𝜃 ∙ 𝐶𝑘
𝜃) ∙ Σ𝜃,𝑘− (63)
Since the weight filter and the state filter have identical output equations, the filters will also share the
same equation for the error. The structure of the DEKF algorithm may be seen as a state EKF inside of
a weight EKF as the weight filter time update is the first step in a given iteration and the weight filter
measurement update the last. The DEKF algorithm is summarized graphically below.
63
Figure 34: The DEKF algorithm. Solid lines represent the paths of state- and weight vectors while dashed lines are the flows of the covariance matrices [125].
Since two filters are run in parallel, the computational strain of the DEKF might be quite large,
depending on the size of state- and weight vectors and the complexity of state- and output equations.
In an effort to speed up computational performance, two different methods of estimating just capacity
degradation are suggested. The first of these, the enhanced state vector extended Kalman filter is, as
the name suggests, based on the EKF algorithm.
3.3.4 Enhanced state vector extended Kalman filter
The ESVEKF is a simplified algorithm for the estimation of cell capacity degradation. The reason for this
simplification is to achieve faster computational performance while still maintaining acceptable SoC
and capacity estimation. The ESVEKF is a variant of the EKF, meaning only one Kalman filter is utilized.
The state vector used in the EKF above has been extended to also feature the cell capacity:
𝑥𝑘 = [𝑉1 𝑆𝑂𝐶 1/𝐶] (64)
This is often called joint estimation, as opposed to the dual estimation of the DEKF and the SWDEKF
which will be outlined below [149]. As the cell SoC is of course dependent on the cell capacity, the A-,
B-, and C matrices of the EKF need to be adjusted accordingly. For the ESVEKF the A-, B- and C matrices
will be the following:
�̂�𝑘 = [𝑒
−∆𝑡𝑅1∙𝐶1 0 00 1 −∆𝑡 ∙ 𝐼𝑘0 0 1
] (65)
�̂�𝑘 = [𝑅1 ∙ (1 − 𝑒
−∆𝑡𝑅1∙𝐶1)
00
] (66)
�̂�𝑘 = [−1𝛿(𝑂𝐶𝑉)
𝛿(𝑆𝑂𝐶)
𝛿(𝑂𝐶𝑉)
𝛿(1𝐶)
] (67)
64
The overall algorithm will then be identical to that of the normal EKF as described above. The partial
derivative of the OCV with regards to 1/C is assumed to be equal to zero, thus echoing the earlier
assumption of the OCV-SoC relation being independent of cell aging.
Note that the ESVEKF will not involve the use of recursive derivatives as the DEKF and SWEDEKF. This
is because the derivative with regards to 1/C is taken with all other elements of the state vector, that
is SoC and V1, as fixed. In the dual estimation approaches, the total derivative, with the accompanying
recursive procedure, is needed as the elements of the state vector are functions of the elements of
the weight vector. This lack of recursive derivatization means the fewer computations must be
performed in each iteration of the joint estimation algorithm. Some authors have claimed that this lack
of recursive derivation might lead to unstable filter behavior however [149]. Another issue might be a
risk of poor numeric conditioning due to the very different rates of change of the elements in the state
vector [150]. Advantages of joint estimation approaches over dual estimation should be the inclusion
of any correlations between changes to especially SoC and capacity due to the common covariance
matrix and the reduced number of internal filter parameters which should be conducive to easier filter
tuning.
3.3.5 Single weight dual extended Kalman filter
The idea of the SWDEKF is to introduce a simple method for estimation of SoH with less computational
strain than the full DEKF. By reducing the weight vector of the DEKF to just a single element, the cell
capacity, the need for heavy matrix operations is dramatically reduced. The SWDEKF is otherwise
identical to the DEKF in terms of overall algorithm as well as state- and output equations, bar the
concatenated weight filter and accompanying matrices. This includes the need for recursive derivation
in every iteration of the algorithm. Another advantage of reducing the size of the weight vector is that
fewer internal filter parameter values need to be specified during the initialization step, meaning that
filter tuning is simplified compared to for the full DEKF. The main concern of this method in comparison
to the full DEKF is the lack of adjustments to the internal resistance parameters which should result in
the filter losing accuracy over time.
In comparison to the other simplified capacity estimator, the ESVEKF, the number of tuning
parameters is larger for the SWDEKF. The main advantage of the SWDEKF over the ESVEKF should in
theory be the faster computational performance as the strain of the many matrix operations of the
Kalman filter algorithm is reduced by scaling down matrix sizes [150]. This is so despite doubling the
number of filters since the execution time of matrix multiplication- and inversion scales to the power
of roughly three with the matrix size. That is for two nxn matrices, the execution time of their
multiplication will scale as n3 [130, 151]. This means that running two filters in parallel, of course these
are interconnected to an extent as well, of sizes two and one, under the assumption that the matrix
operations are the limiting factor in terms of performance, is more effective than running one filter of
size three as 33=27>(23+13)=8. In practical use, there are of course additional factors to be considered
than just matrix operations and measured actual computational times will therefore be compared.
65
4 RESULTS AND DISCUSSION
4.1 EXTENDED KALMAN FILTER As the most basic variant of the Kalman filter, the LKF, is in essence not applicable for the Li-ion battery
monitoring due to the strong non-linearity of the system, it was for the linearized EKF discarded in this
work. The EKF, still the most basic algorithm investigated in this study, requires the allocation of four
values of internal filter parameter in the initialization stage: the starting guesses of the state vector
and state covariance matrix, the process noise covariance matrix, and the measurement noise
covariance matrix. The effects of all these parameters on filter performance will now be presented and
discussed. What will become evident is that while the EKF certainly is a very powerful SoC estimator,
it does have its limitations and pitfalls. Fortunately, after a careful analysis of the filter behavior these
issues can be avoided relatively easily or at least managed effectively. The EKF proves successful in
overcoming most the issues of the Coulomb counting method, as discussed in section 1.4, including
sensitivity to starting SoC value, and sensitivity to sensor noise. The EKF is also found to outperform
Coulomb counting in terms of SoC estimation accuracy for cases of inaccurately specified cell
capacities.
As the EKF, here being used purely as an SoC estimator, does not update any of the values of ECM
parameters, the initial values of these were used unless otherwise stated. These values were
determined using battery diagnostic test before the first RW phase as described earlier in section 3.1.
The reference method for validation of the SoC estimation results was Coulomb counting for all SoC
work. Coulomb counting was applicable as a reference method as the batteries were always fully
charged before a RW phase began, meaning that the starting SoC was exactly known, and as current
and time were measured with high precision laboratory equipment, thus minimizing sensor bias and
other measurement error. It should however be noted that the quality of the Coulomb counting
reference will degrade over time as the cell capacity used for its calculation will be assumed to be
constant while the actual cell capacity will change due to variations in temperature, current and most
importantly due to aging. Assuming that capacity degradation due to aging is the dominating effect,
the Coulomb counting should be at its most accurate at the beginning of each RW phase. Therefore,
mainly these operation periods were used for validation and comparison of SoC estimation results.
As Coulomb counting is already indirectly included in the model equations of the applied ECM, more
specifically in the state equation, see (39), the SoC estimation of the filter should follow the Coulomb
counting reference very closely when the starting value of SoC specified during initialization of the
algorithm is correct, in this case SoC0=1 as the battery is known to be fully charged at the start of any
given RW phase. This scenario, with the starting guess for the transient voltage being zero, is shown in
figure 35 below over the first 100 RW steps of the first RW phase of battery RW9. The internal filter
parameters used for obtaining the results in the two figures below are summarized in table 9.
66
Table 9: Internal EKF parameters for the results in figures 35 and 36.
Internal filter parameters
Starting SoC 1.00
Starting V1 0.00 V
Covariance, SoC 1∙10-2
Covariance, V1 1∙10-3 V2
Process noise, SoC 1∙10-10
Process noise, V1 1∙10-4 V2
Measurement noise 5∙10-3
Figure 35: SoC estimated by EKF over first 100 RW steps of battery RW9.
Looking at figure 35, it is evident that the Kalman filter SoC estimation indeed follows the Coulomb
counting reference very closely in this case. As none of the common issues of Coulomb counting appear
in the test above no advantages of the EKF are manifested either. If one examines a later stage of the
same RW phase, such as in figure 36 below, one will notice a larger difference between the EKF SoC
estimate and the Coulomb counting reference. The reason for this is partly that modelling errors grow
larger over time due to the ECM parameter values not being adjusted to compensate for aging effects.
This means that the state covariance matrix will get bigger (individual elements will assume larger
values) leading to an increased state uncertainty and thus variance in SoC estimation. The Coulomb
counting reference will also be less accurate towards the end of any given RW phase as the real cell
capacity has been lowered from the initial measured value due to aging effects. Since one can be sure
that the Coulomb counting reference is inaccurate at this stage, but not to what extent, it is difficult to
quantify the error of the EKF SoC estimation too. While it is impossible to verify SoC estimation results
in this case of no valid reference it should be noted that the EKF and other Kalman based approaches
have the potential of being more accurate than Coulomb counting in cases of uncertain cell capacity.
This is due to the state estimate measurement update (29) adjusting the SoC estimate to match the
measured cell voltage. The state estimate time update equation, keeping in mind the state equation
(39) of the utilized ECM, is identical to Coulomb counting, but the additional inclusion of the cell voltage
measurement in the correction stage of the EKF algorithm should allow for a more accurate SoC
estimation than when solely relying on Coulomb counting.
67
Figure 36: SoC estimated EKF during end of first RW phase of battery RW9.
If one keeps the same ECM parameter values, that is values of R0, R1, C1 and cell capacity, for even
longer, the resulting EKF SoC estimation error is more noticeable. The figure below shows an example
simulationfrom the 5th RW phase of RW9, recorded a little more than three weeks into testing of RW9.
Since figure 37 shows the start of a RW phase, the Coulomb counting reference may be regarded as
being accurate as the capacity used for its calculation has been adjusted to the latest measured value.
SoC estimation accuracy has clearly degraded compared to the results above. The need for online ECM
parameter adjustment to achieve consistent acceptable SoC estimation accuracy over time is evident,
as are the limitations of the EKF algorithm. Still, the SoC estimation results are not too far of the
Coulomb counting reference values despite the now quite extensive modeling error. The root-mean-
square error of the EKF SoC estimate compared to the Coulomb counting reference is 4.3 % in figure
37. As a comparison, if normal Coulomb counting would have been applied under the same
circumstances, that is with an unadjusted capacity, the resulting root-mean-square error to the
Coulomb counting with adjusted capacity would be a whole 10 %. This then confirmed the discussion
above regarding the EKF being more accurate than Coulomb counting for cases of an inaccurate cell
capacity. This is of course a great advantage of the EKF over Coulomb counting and effectively results
in capacity recalibration, usually meaning a total discharge of the battery from the fully charged state,
being necessary less frequently. The issue of negative SoCs discussed briefly in section 3.2 is also
evident in figure 37.
68
Figure 37: SoC estimation by EKF with unadjusted ECM parameters for RW phase 5, battery RW9.
After having seen two cases of the EKF not performing noticeably better than (capacity-adjusted)
Coulomb counting, some scenarios which highlight its upsides will now be considered. One of the most
significant strengths of the EKF, especially compared to Coulomb counting, is that the initial SoC does
not need to be known exactly for SoC estimation to be successful. The filter should be able to converge
to the true SoC within a reasonable time given a fairly accurate battery model, at least if values of
internal filter parameters have been sensibly chosen to reflect the uncertainty of the initial state. In
fact, as seen in the next figure below, the initial guess of SoC could indeed be very poor and the EKF
would still find the correct value after an initial period of adjustment. The interaction between the
initial guesses of the state vector and the state covariance matrix was found to be critical to this
behavior and will now be analyzed.
Cell SoC must in theory always lie in the range between zero and one (0 % - 100 %). If the initial SoC
should be totally unknown, the average starting SoC guess error would thus be 0.5 on average. Such a
situation seems unlikely as one will usually at least have some vague idea of how charged the battery
is, meaning that the real average initial error is likely smaller. The initial guess of transient voltage is
easier to determine as this will almost always be zero, assuming that estimation starts when operation
of the battery starts. This is because the transient voltage will naturally decay to zero over time when
no current is being applied. For the analysis of the impact of the initial state vector guess on SoC
estimation accuracy, only the initial SoC will therefore be varied in the analysis below. First only the
initial guess of SoC was varied, meaning that all other internal filter parameters of the EKF were set as
in the previous two figures. In the figure below the starting SoC guess has been set to 0.75, 0.5, and
0.05 respectively.
69
Figure 38: EKF converging to correct SoC value despite incorrect starting values.
As can be seen in figure 38, the EKF will converge to the “true” value represented by the Coulomb
counting reference despite even very poor initial guesses of SoC, thus seeming to be quite robust to
poor initialization. The difference when the starting guess is poor is that it takes the filter a longer time
to converge to the Coulomb counting reference, stabilizing after approximately 0.03 hours (less than
2 minutes) of operation for a starting SoC guess of 0.75 and after approximately 2 hours for a starting
SoC guess of 0.05 with this set of internal filter parameters. After the filter SoC estimate has converged,
the estimation is of equal accuracy for all starting guesses of SoC and virtually identical to the Coulomb
counting reference.
It was found that the rate of estimation convergence to the Coulomb counting reference was also
dependent on the initial SoC covariance. The case of a constant starting SoC guess of 0.5 with a varying
initial SoC covariance was considered below. Since the real SoC starting value was known, an
appropriate starting value for the SoC covariance could easily be determined using the rule of thumb
in (20): Σ𝑥,0 = (1-0.5)∙(1-0.5) = 0.25. The values for the starting guesses for SoC covariance have been
set to 0.25, 0.01, and 0.001 in the figure below, all other internal filter- and ECM parameters being set
as in table 9.
70
Figure 39: Differences in convergence rate to Coulomb counting reference value for varying starting SoC covariance.
Here it is interesting to note that convergence is fastest for the estimate with the middle value
Σ𝑥,0=0.01, which approached the Coulomb counting reference almost immediately, while the lower
starting covariance value results in the estimated SoC reaching the Coulomb counting reference after
around two hours. It is also evident that when the starting SoC covariance is set to 0.25 there is not
enough time for the filter estimate to converge within the first 100 RW step at all. If the simulation
was run for a longer period, convergence would however occur eventually, as seen below.
Figure 40: Eventual convergence of EKF SoC estimate despite poor choice of initial SoC covariance.
Note the values on the time axis in figure 40. The time needed until convergence had occurred could
be very long for large values of initial SoC covariance. If the convergence behavior is this slow, issues
might arise due to substantial changes to the values of ECM parameters due to aging having occurred
71
before converging. If this is the case, the filter may diverge and never find the true SoC value at all. It
bears repeating that the mode of cycling used for battery RW9 is very intense and aging effects are
thus accelerated compared to during most normal Li-ion battery applications.
The effect of changing the starting state covariance matrix was evidently found to be quite large. The
optimal value of the initial SoC covariance also appeared to be lower than 0.25, which puts the validity
of the rule-of-thumb method in (20) into question. It was also notable that the filter would still
converge to the Coulomb counting reference eventually, even for seemingly too high initial values of
the SoC covariance, just that it would take a very long time in some cases. On the other hand, if the
initial covariance was set much too large there was a risk that the filter would diverge completely. The
idea when specifying the initial state covariance was then found to be something like this: specify a
large initial covariance if the initial state uncertainty is large to allow for fast convergence, however
risking low stability (risk of estimation diverging) in the process. On the other hand, setting a small
initial covariance lessens the issues of estimation instability but might instead result in extremely slow
convergence [63]. Again, a worst-case scenario in the case of Li-ion battery monitoring is that the
convergence is slower that the rate of change of ECM parameter values which could lead to total
divergence. Finding appropriate values of the initial state covariance matrix therefore becomes
something of a balancing act, especially for more uncertain starting states.
Finally, the unvarying internal EKF parameters were examined: the process- and measurement noise
covariance matrices. Simply put, these parameters decide the “trust” (or “distrust”) of the
measurements relative to the model predictions. First the effects of changing the value of the
measurement noise parameter were investigated. To show this in the clearest way possible some
artificial noise is added to the cell voltage signal. This is done to simulate a very noisy, worst-case-
scenario measurement signal as a kind of stress test of the algorithm. The added noise is Gaussian
(normally distributed) and centered around the actual measured value. In other words, the accuracy
of the cell voltage measurement was still the same as before but its precision has been artificially
lowered. The easiest way of accomplishing this in MATLAB was to multiply each element in the vector
of measured cell voltages by a normally distributed number centered on 1 with a standard variation of
b in this way:
𝑉𝑡,𝑛𝑜𝑖𝑠𝑦 = 𝑉𝑡.∗ (1 + 𝑏 ∙ 𝑟𝑎𝑛𝑑𝑛(1, 𝑙𝑒𝑛𝑔𝑡ℎ(𝑉𝑡)) (68)
The MATLAB function “randn” being a generator of normally distributed random numbers centered
on zero. For these test the standard deviation of the noise was set to 5∙10-3 V, producing a suitably
noisy voltage signal. For the three simulations in the figure below, the measurement noise covariances
were set to 0.5, 5∙10-3 and 5∙10-5 V, respectively. The variation of the measurement noise covariance
between these simulations was quite massive in other words (four powers of ten). One of the main
advantages of Kalman filter based approaches is their ability to deal with noisy input- and output
signals. By this it is meant that the algorithm can separate out the true signal from noise and thus
achieve a better state estimate than what would be possible using a deterministic approach such as
Coulomb counting.
72
Figure 41: EKF SoC estimation with different measurement noise covariances.
The starting SoC was set to 0.5 for all tests shown in figure 41. The effect of changing the measurement
noise covariance on the convergence rate of SoC estimation can be seen to be somewhat like that of
changing the initial state covariance matrix. This is sensible when considering (24). The calculation of
the Kalman gain matrix controls how heavily state estimations should be adjusted. If the measurement
signal is known to be very noisy, meaning that a high measurement noise covariance has been
specified, then the model prediction should hold more weight and vice versa. While this means more
accurate SoC estimation overall, it might also mean slower convergence. This effect is observed in
figure 41. If one eliminates the effect of slower convergence by specifying a correct starting SoC, then
the general a trend can be seen between the logarithm of the measurement noise covariance and the
root-mean-square SoC estimation error:
Figure 42: Overall trend of rms SoC error versus logarithm of measurement noise covariance for cases of noisy signals and correct starting SoC for the EKF.
73
The results in figure 42 shows the average rms SoC error versus the Coulomb counting reference over
five simulations of the 300 first RW steps of the first RW phase of battery RW9. All internal filter
parameters were as in table 9 with standard deviations of current- and voltage signal noise equal to
values in figure 41. It is clear that the SoC estimate accuracy increases with increasing measurement
noise covariance for very noisy measurement signals if convergence of the estimate does not need to
be considered. Of course, if measurement noise covariance is increased even further, the EKF will
essentially turn into Coulomb counting as the cell voltage signal is eventually completely ignored,
resulting in an rms error of zero in the case shown in figure 42 as Coulomb counting is the reference
method.
Of course, in this work where carefully recorded laboratory measurements were used, the
measurement noise was generally very small, meaning that the interpretation of the cell voltage signal
is straightforward. In a more practical situation this might not be the case, as can be seen in the figure
below where the EKF cell voltage predictions corresponding to the same set of measurement noise
covariances as in figure 41 above are shown.
Figure 43: Cell voltage predictions of the EKF for the case of a noisy voltage signal using different measurement noise covariance matrices.
As can be deduced from the plot of cell voltages above, lowering the measurement noise covariance
might cause the algorithm to interpret measurement noise as true signal. This will lead to continuous
overcompensation of the state estimates, per the measurement update equation (27), reducing the
precision of not only the cell voltage prediction but also the SoC estimation. This overfitting-like
behavior can be seen especially for the case of Σ𝑣=5∙10-5 above. On the other hand, a higher value of
the measurement noise covariance allows for the Kalman filter to better recognize the true cell voltage
signal and thus stability and precision is clearly improved. In figure 43 this can be seen for the
simulation using Σ𝑣=0.5, which follows the true, as measured, voltage signal quite closely despite the
quite heavy added artificial noise. Considering the wide range of values of the measurement noise
covariance in the figure above, one can also draw the conclusion that the SoC estimation is not
especially sensitive to this parameter, even though it does have an impact. Naturally, the importance
of this parameter is heightened for scenarios with higher sensor noise and bias.
74
The process noise covariance parameters can be said have a somewhat similar effect on results as the
measurement noise covariance, controlling how much “trust” one puts in the model versus
measurements. One difference of the process noise covariance matrix to the measurement noise
covariance is that here there are two separate parameters interacting with each other. The relative
magnitude of the process noise covariance of the transient voltage compared to that of the SoC
estimate turned out to be an important factor in achieving overall algorithm precision and stability.
The most accurate SoC estimation results were achieved when the process noise covariance of the
transient voltage was set to a relatively higher value than for the SoC, such as has been the case for all
simulations so far in this section. This is quite sensible when considering how the development of the
ECM was conducted in this work. The state equation in the EKF calculates the SoC using only one
prespecified ECM parameter, the cell capacity. This is a carefully measured value, using what is
essentially Coulomb counting and an accurate 13th degree polynomial OCV-SoC expression. On the
other hand, the estimation of the transient voltage V1 uses the estimated values of both R1 and C1. Due
to the high complexity of battery transient behavior most of the modelling error of a given ECM will lie
here. This is the reason why more advanced ECMs are generally expanded by adding on more resistors
and capacitors, while the treatment of the OCV-SoC relation and the capacity determination largely
remains the same [104]. When one compares the error in estimation of the OCV-SoC relation in figure
24 with the overall modeling error in 28 for instance, it is clear that much of modeling error stems from
the transient behavior of the cell. Thus, the transient voltage estimate should logically be less accurate
than that of SoC, which is reflected in the appropriate values in the process noise covariance matrix.
This principle of “balancing” modeling error will be seen to be an even more important aspect for SoH
estimation. Finally, some practical usability aspects of the EKF will now be discussed.
In the case of the data used in this work, which is recorded from laboratory experiments, Coulomb
counting is a valid SoC reference method for the reasons discussed at the start of this section. In a
more realistic situation, where no reasonable reference for the SoC estimation exists, one will know
that the filter has converged by gleaning at the values in the state covariance matrix. diagonal values
of the state covariance matrix should be continuously decreasing, if internal filter parameters were set
suitably anyway, and should eventually stabilize to a low value if the filter has converged, given that
the accuracy of the applied battery model is reasonable. This will be demonstrated below by
application of the rule-of-thumb in (21).
As mentioned earlier, one of the inherent strengths of Kalman filter based approaches to state
estimation are the “built-in” error bounds. While (21) strictly only applies to the linear Kalman filter, it
may still be used for more advanced Kalman filters to get a rough idea of the estimation certainty. In
the figure below upper- and lower error bounds, calculated using (21), are plotted for SoC estimation
using the EKF with noisy signals of voltage and current. Added artificial noise was on the form of (68)
and had a standard deviation of 5∙10-3 V and 5∙10-3 A. The starting guess of SoC was 0.5.
75
Figure 44: Error bounds per (20) of SoC estimate using EKF.
It is clearly visible in figure 44 how the error bounds get closer to the estimate and the Coulomb
counting reference value as filtering goes on, meaning that the SoC estimate certainty increases. It
intuitively seems sensible that as the estimation certainty should increase with increasing information,
in this case measurements of current and cell voltage, becoming available. This principle is reflected in
(30) of the measurement update equations. After this initial period of the error bounds getting closer
to the SoC estimate, the SoC estimate covariance should stabilize and remain roughly constant for the
remainder of filtering unless modelling- or measurement errors change significantly in magnitude. This
is shown in the figure below. Monitoring of the diagonal values of the state covariance matrix therefore
becomes a useful tool in ensuring Kalman filter performance.
Figure 45: The stabilization of SoC covariance in the EKF, indicating estimation convergence.
76
4.2 ADAPTIVE EXTENDED KALMAN FILTER The AEKF has the distinct advantage of updating process- and measurement noise covariance matrices
during operation of the filter. This would in theory mean that the filter tuning process could be made
simpler and also that the internal filter parameters would automatically be updated if operation
conditions were to change in such a way as to affect the measurement- and/or process noise [130].
The algorithm is as mentioned earlier very similar to the EFK in most aspects and indeed produced very
similar, if not even better, results in terms of accuracy and convergence rate of SoC estimation. The
ECM parameter values obtained from pulsed- and reference discharge tests, as presented in tables 7
and 8 respectively, were used to obtain all AEKF results. Below the AEKF SoC estimation over the 100
first RW steps of the first RW phase of RW9 is shown against the Coulomb counting reference.
Figure 46: SoC estimation for 100 first RW steps using AEKF with Coulomb counting reference.
Note the very poor starting guess of SoC0=0.05 in figure 46. The AEKF SoC estimate converges to the
values measured with Coulomb counting extremely quickly and thus appears quite robust to poor
initialization. It is of course interesting to compare these results with the results for the same starting
SoC for the EKF in figure 38. The convergence rate of the AEKF SoC estimate is noticeably faster than
for the EKF. The internal filter parameters used in obtaining the results above are summarized in table
10.
Table 10: Internal filter parameters for AEKF for figure 46.
Internal filter parameters
Starting SoC 0.05
Starting V1 0.00 V
Covariance, SoC 1∙10-2
Covariance, V1 1∙10-3 V2
Starting process noise, SoC 1∙10-2
Starting process noise, V1 1∙10-2 V2
Starting measurement noise 1∙10-2
Moving estimation window size, M 10
77
As can be seen in table 10, the starting state covariance matrix was the same as for the results obtained
with the EKF above. Note the very high values used for starting guesses of the process noise covariance
matrix and the measurement noise covariance. These starting values could furthermore be varied
substantially without affecting estimation accuracy considerably. To demonstrate this behavior, the
initial process noise covariance matrix was varied by multiplying the starting matrix values in table 10
by 0.01 and 100, respectively. All other internal filter parameters were as in table 10. As can be seen,
convergence remains very rapid for all selected initialization values:
Figure 47: Rapid convergence of AEKF SoC estimate for various initial process noise covariance matrices.
Note the values on the axes in the figure above, the scale of the time axis has been made much smaller
compared to in figure 46. As can be seen, larger values in the initial process noise covariance matrix
seem to indicate faster convergence to the Coulomb counting reference. The variation of the initial
process noise covariance matrix is quite large in figure 47, indicating that while its setting does matter,
its effect on results was overall found to be very small. A starting process noise covariance larger than
one is of course most often senseless in the context of SoC estimation. Assignment of very small values
in the starting process noise covariance matrix could lead to the AEKF SoC estimate diverging for lower
values in the initial state covariance matrix and should therefore be avoided.
If one instead varies the starting measurement noise covariance a similar conclusion is reached, as is
seen below.
78
Figure 48: The quick converging of the AEKF for various starting measurement covariances.
Again, the scale of the time axis in figure 48 is notable. While the difference in results for different
starting measurement covariances is relatively slim, it is seen in figure 48 that lower initial
measurement noise covariance led to slightly faster convergence rates in this case. Note also the large
range of values of the initial measurement noise covariance in figure 48, covering a full six powers of
ten. The specification of this internal filter parameter was in other words found to be virtually arbitrary
to overall filter performance in terms of accuracy and convergence rate.
The final new parameter of the AEKF compared to the EKF is the moving estimation window size, M,
as in (42). The effects of changing this parameter were once again overall very modest. The biggest
influence was seen when artificial noise was added to the measurement signals of cell voltage and
current. For the results below the artificial noise was added on the form (68) with a standard deviation
of 5∙10-3 V and 5∙10-3 A, respectively. All internal filter parameters except the moving estimation
window size were as in table 10.
Table 11: The effect of moving estimation window size on AEKF SoC estimation accuracy.
Moving estimation window size, M Root-mean-square SoC error compared to Coulomb counting over 100 first RW steps
5 0.0149
100 0.0199
200 0.0222
It seems then that a smaller moving estimation window size is preferable from an estimation accuracy
standpoint. Finding literature on reasoning of the choice of moving estimation window size seems
difficult, Campestrini et al. used a moving estimation window size of ten in their recent review
seemingly without motivating the choice [130]. Since this value seemed to produce acceptable results
in that work and likewise when applied here it was used throughout. Sun et al. instead found a moving
estimation window size of 100 to produce the best results for use in the AEKF, thus seemingly
contradicting the results observed here [117]. Looking at table 11, the moving estimation window size
appears to be very much a fine-tuning parameter for the case of Li-ion battery monitoring. This seems
79
to confirm the conclusion in [145] that the moving window size has little influence on overall filter
performance.
The observed effects of changing the initial state covariance matrix of the AEKF was found to be
identical to those observed for the EKF and will therefore not be discussed further. Finally, for the case
of SoC estimation given a faulty assignment of cell capacity, as seen in figure 37 for the EKF, the AEKF
was found to perform slightly better than the EKF, achieving a root-mean-square error of 4.2 % in the
identical situation as discussed in section 4.1.
4.3 DUAL EXTENDED KALMAN FILTER Having explored two different Kalman filters for SoC estimation, the focus turned towards methods
for estimation of SoH. First the DEKF was investigated. This method utilizes two extended Kalman filters
working in parallel, one for state estimation (e.g. SoC) and one for weight estimation (e.g. capacity).
There is also some information exchange between the filters. Out of the three examined SoH
estimators, the DEKF was the only one capable of estimating not only the cell capacity degradation,
but also the increase of internal resistance over time. For this reason, it is very interesting to compare
results of the DEKF with results for the other two SoH estimation algorithms which are only adjusting
the capacity during filtering. For validation of all capacity estimation, the measured cell capacities in
reference discharge tests, as described in section 3.1.1, were used as reference values.
The figure below show a plot of cell capacity estimation using the DEKF over the first full RW phase of
RW9. Table 12 summarizes all the internal filter parameters that were used in initialization of the DEKF
in figure 49. As can be seen, the capacity estimate converges relatively quickly to a value close to the
measured capacity before the first RW phase, and then slowly decreases to eventually reach a value
close to the measured value before the second RW phase by the end of the simulation, see table 7 for
these measured capacities. Here the initial cell capacity value was set to 1.9 Ah. The initial guess of SoC
was 0.999, and was thus taken as wholly accurate considering the fully charged state of the cell before
each RW phase. It is believed that this corresponds to a realistic scenario since fully charging a cell
using a constant current-constant voltage (CC-CV) scheme is always possible if the upper cut-off
voltage of the battery is known. As such, the corresponding SoC initial covariance has been set to a low
value, reflecting the relative certainty of the initial SoC guess.
After the initial converging mode during the first four hours or so of filtering, the DEKF estimates an
approximately linear loss of capacity over time which corresponds well to data in the literature [88,
152, 153], keeping in mind the relatively unique mode of operation used for obtaining the data used
in the current study. This linear capacity loss over time also seems within reason when considering the
trend of measured capacities in reference discharge tests over the first twelve RW phases of RW9 as
seen in figure 21. Achieving reasonable results with the DEKF required careful consideration of the
internal filter parameters, filter tuning that is. This will be discussed further below.
80
Figure 49: Capacity estimation over first full RW phase of RW9.
The measured capacities via reference discharge tests were 2.10 Ah before and 2.04 Ah after the RW
phase simulated in figure 49. It is difficult to exactly quantify the accuracy of the estimate of the initial
capacity here due to the initial period of estimate convergence, but the highest capacity estimated by
the DEKF in figure 49 is 2.11 Ah, which would be equivalent to a relative error of less than 0.5 %. The
capacity estimated by the DEKF algorithm at the end of the RW phase above is 2.02 Ah, thus
corresponding to a relative error of less than 1 %. The possibility of estimating the battery SoH in terms
of capacity fade using the DEKF therefore seems promising.
Table 12: Internal filter parameter values for initialization of DEKF in figure 49.
State filter internal parameters
Starting SoC 0.999
Starting V1 0.0 V
Covariance, SoC 1∙10-10
Covariance, V1 1∙10-5 V2
Process noise, SoC 1∙10-10
Process noise, V1 1∙10-8 V2
Measurement noise, state filter 5∙10-3 V2
Weight filter internal parameters
Starting capacity 1.9 Ah
Starting R0 0.070 Ω
Starting R1 0.032 Ω
Covariance, capacity 1∙10-4 (Ah)2
Covariance, R0 1∙10-9 Ω2
Covariance, R1 1∙10-9 Ω2
Process noise, capacity 1∙10-11 (Ah)2
Process noise, R0 1∙10-11 Ω2
Process noise, R1 1∙10-11 Ω2
Measurement noise, weight filter 1∙10-2 V2
81
If the initial SoC would also be unknown, the initial SoC covariance must be increased to reflect this. In
the figure below the starting SoC was set to 0.5 with an initial SoC covariance of 1∙10-4. Note how the
filter will find the correct capacity and SoC concurrently despite poor initial values of both. It should
be noted that the specification of initial covariance matrices was found to become more critical, and
sensitive, when starting values of several closely connected parameters, such as the SoC and cell
capacity, were uncertain. It was very easily so that one of the two connected parameters, such as SoC
and capacity, would be over-adjusted if initial covariance matrices were not chosen carefully. This
could lead to the filter diverging. This is believed to be due to a certain “entanglement” effect of the
parameters due to the two filters sharing the same output equation. This will be discussed further
below.
Figure 50: SoC- (left) and capacity (right) estimates by DEKF when initial guesses of both SoC and capacity are poor for 100 first RW steps of first RW phase of RW9.
Note in figure 50 above how the SoC estimate converges to the Coulomb counting reference
approximately simultaneously as the capacity estimate reaches a value close to what was measured
before the RW phase, in this case 2.10 Ah.
The DEKF was also found to be able to be set up to quickly estimate the internal resistance parameters,
given unknown or inaccurate starting values of these. The initial guess of the cell capacity was then set
to the value measured before the first RW phase, 2.10 Ah, and the related initial capacity covariance
was correspondingly set to a low value of 1∙10-7 to separate out the effects of capacity estimation for
these tests. The internal filter parameters relevant to estimation of internal resistance parameters
were set as in table 13:
Table 13: Internal filter parameters of DEKF for results in figure 51.
Starting R0 0.050 Ω
Starting R1 0.070 Ω
Covariance, R0 1∙10-5 Ω2
Covariance, R1 5∙10-6 Ω2
Process noise, R0 1∙10-11 Ω2
Process noise, R1 1∙10-11 Ω2
82
Figure 51: Rapid convergence of both internal resistance parameter estimates using the DEKF despite poor initial guesses.
As can be seen in figure 51 both internal resistance parameter estimates would converge and stabilize
despite relatively poor initial guesses of both, the one for R1 being too high and one for R0 being too
low in the case shown. When comparing to measured values of internal resistance parameters in table
8 it may be seen that both resistance estimates are too low in figure 51, R0 by 7 % and R1 by 13 %,
approximately. This might be due to temperature effects which have not been accounted for in the
current model. It is widely known that the internal resistance of a Li-ion cell will be lower at higher
temperatures. In this case, the average temperature during the pulsed discharge test preceding the
first RW phase of RW9 was 23.4 ⁰C while it was 29.8 ⁰C during cycling. The average temperature was
in fact lower for all pulsed discharge tests in comparison to the subsequent RW phases, as seen in table
14.
Table 14: Differences in temperature between pulsed discharge tests and RW phases for RW9.
Before RW phase/during RW phase
Average temp., pulsed discharge test [⁰C]
Average temp., RW phase [⁰C]
Difference [⁰C]
1 23.4 29.8 6.4
3 23.8 29.0 5.2
5 23.3 28.7 5.4
7 29.0 35.5 6.5
9 30.0 35.2 5.2
11 30.1 34.3 4.2
13 28.9 34.2 5.3
Since the temperature dependence of the internal resistance parameters can be quite large, and was
not determinable for the used data set, there is reason to believe that the filter is converging to the
correct values despite the differences to measured values [127, 154]. While unfortunately there is no
way of absolutely validating these estimates of internal resistance, some weight must be put on the
ability of the algorithm to accurately predict the SoC and capacity using these values. In fact, the values
of internal resistance estimated by the DEKF were found to consistently produce better results than
those from the pulsed discharge tests when inserted into the EKF algorithm in terms of root-mean-
square error of SoC estimation.
83
To clearly illustrate the advantages of the DEKF over the EKF for long-term SoC estimation another set
of operational data from another battery was used. While still belonging to the “Randomized Battery
Usage Data Set”, the Li-ion battery RW2 was operated quite differently compared to RW9 during the
RW phases. Here it was the duration of charging that was randomized from the set [0.5 hours, 1 hour,
1.5 hours, 2 hours, 2.5 hours] and the charging current was instead always 2 A. The battery was charged
using this current for the randomized duration or until the cell voltage reached the upper cut-off
voltage 4.2 V. At this point, the charging changed to a constant voltage mode which continued until
the applied current dropped below 0.01 A or the randomly chosen duration had ended. This is then
essentially CC-CV charging, meaning that the battery is periodically fully charged and thus often
reaches points of 100 % SoC. Discharging was done as with RW9, meaning that a random current
between 0.5 A and 4 A is applied for five minutes or until the lower cut-off voltage is reached. A plot
of a series of charge- and discharge cycles in the form of measured cell voltage and current from this
data is shown in the two figures below.
Figure 52: Measured cell voltage for RW cycling using CC-CV mode charging for RW2. Note the periods of constant voltage at 4.2 V.
The points in figure 53 below were the current approaches zero will roughly represent points of 100 %
SoC.
Figure 53: Current over time for cycling of battery RW2.
84
Below the SoC estimated by the DEKF for this data is shown, along with a Coulomb counting reference
which uses the capacity measured in reference discharge tests before cycling began.
Figure 54: SoC by Coulomb counting and DEKF for CC-CV mode charge cycling of RW2.
Due to the existence of these points of almost guaranteed 100 % SoC it is easy to see in figure 54 that
the DEKF performs rather well in terms of predicting the SoC and thus cell capacity. The Coulomb
counting, which of course does not adjust for changes to cell capacity, quickly loses its accuracy and
even predicts an SoC below zero at certain points in figure 54. Of course, like Coulomb counting the
EKF does not adjust for changes to cell capacity either and would thus struggle to perform much better
in this scenario. It is also clear that the capacity fade is very rapid at these cycling conditions of high
ΔSoC, as was expected from the discussion of cycle aging impact factors in section 1.2.4.2. In addition,
some other advantages of the DEKF became evident with this test. Firstly, no data from reference
discharge tests needed to be considered for determination of initial cell capacity as a rough initial cell
capacity could be guessed from the earlier analysis of the similar battery RW9 and this estimate was
then adjusted by the DEKF online to fit the data. Secondly, no pulsed discharge tests for determination
of the values of internal resistance parameter values were needed as, again, rough guesses using
experience from battery RW9 were sufficient. The only battery-specific data that was required for the
initialization of the DEKF for the “new” battery was low current discharge data for determination of
the OCV-SoC relationship of the cell, which was fitted with a 13th degree polynomial just as with RW9.
In other words, using the DEKF instead of the EKF could potentially reduce the need for likely time-
consuming and expensive cell-characterization tests, at least if rough estimates of initial cell
parameters are available. Another advantage of using the cycling data from RW2 for algorithm testing
and validation is that points of negative SoC, as mentioned in section 3.2, are avoided. While it is
difficult to quantify the effect that these points had on results, it does appear that especially estimates
of capacity are more stable when using the data from RW2. For instance, it was found that the capacity
estimate was less dependent on the initial guessed value and would eventually converge to a common
value independent of the starting guesses more often. This is shown in the figure below where starting
guesses of cell capacity were 1.7 Ah, 1.9 Ah and 2.2 Ah respectively.
85
Figure 55: Convergence of capacity estimates despite large disparity in starting values using DEKF and data from battery RW2.
Note the relatively turbulent initial capacity estimates, followed by stabilization and convergence to a
common value. The line representing measured capacity in figure 55 shows the capacity measured in
reference discharge tests before cycling began. The cell capacity measured after the cycling in figure
55 was 1.92 Ah, thus closely matching the estimated value. The relatively slow rate of convergence is
also notable. For the data acquired from RW9, the capacity estimates tended to diverge if the
convergence was not very quick. The fact that such slow convergence as seen in figure 55 was at all
possible indicates a certain robustness of the capacity estimation. Certainly, convergence could also
be made quicker by increasing the initial capacity covariance, as seen below.
Figure 56: Quick convergence of capacity estimates with different initial guesses of capacity for DEKF using battery data from RW2.
86
Note the difference in the scale of the time axes in figures 55 and 56. As can be seen to the left of
figure 56, higher values of initial capacity covariance do seem to cause an increased initial estimate
turbulence, which might be an issue. All internal filter parameters were set the same in obtaining the
results shown in the two figures above, bar the initial capacity covariance which was set to 10-6 (Ah)2
and 10-3 (Ah)2 in figure 55 and 56 respectively.
As is evident per table 12, the number of internal filter parameters that must be set for initialization
of the DEKF is very large, much larger than for the EKF: 17 for the DEKF compared to just 7 for the EKF.
This is troublesome, especially considering the lack of strict, intuitive guidelines as how to set most of
these, even though some general principles and strategies learned from the EKF tuning may be applied.
Also, the curse of dimensionality means that exploring combinations of these parameters quickly
becomes unmanageable. Of course, the state- and weight filters and their internal filter parameters
will interact with each other as well. This means that also the relative magnitude of parameter values
in the filters must also be considered. The trial-and-error approach of identifying covariance matrix
values outlined in section 3.3 of course becomes exponentially more time consuming with an
increasing number of parameters to be set. Despite these issues, which hindered the full
characterization of DEKF filter tuning behavior, some general trends could be elucidated.
For the initial guesses and initial covariance matrices the same general principle as for the EKF applies,
that is if initial guesses are uncertain or intentionally poor then the initial covariance matrix/matrices
should be set larger to ensure correct convergence. Values of the process- and measurement noise
covariance matrices, which are now two in number respectively, are still the most challenging to
determine and a trial-and-error approach is a must for most cases, including this work. As was seen in
the discussion of the EKF above, by balancing the process- and measurement noise covariance matrices
it was possible to achieve accurate estimation despite very noisy input- and output signals for instance.
In the case of using two filters simultaneously, this procedure becomes more complex. As can be seen
in table 12 and in earlier descriptions of the DEKF algorithm, measurement noise is to be set separately
for the state- and weight filters. This might seem odd considering that the two filters share the same
output equation and, thus, measurement of cell voltage. Much like when considering the very simple
state equation of the weight filter, this is a question of the difference in rates of change of the elements
of the state vector compared to those in the weight vector.
Take the cell capacity as an example. Say that in a certain time step a difference between the predicted-
and measured cell voltage is recorded, an error per (28). The goal is now to adjust first the state vector
and then the weight vector by using this error multiplied by the corresponding Kalman gain matrix per
(29) and (62). Looking at the output equation, this is a summation of three terms: the OCV, which is a
function of the SoC; the transient voltage V1, and the instantaneous ohmic resistance response term
I∙R0. The error in the calculated cell voltage might be due to an error in the estimation of any of these
terms individually, some combination thereof, or all of them simultaneously. For the EKF the situation
is now relatively simple. Since only the SoC and the transient voltage are adjusted on an online basis,
one will simply adjust one or both based on the current state covariance matrix and the C matrix, along
with the specified constant measurement noise covariance. For the DEKF things get more complicated.
A faulty SoC prediction might be due to either the state filter as in the EKF, but might also be caused
by a faulty cell capacity estimate. This is because the SoC is a function of the cell capacity, as is evident
by considering its definition in (9). The same situation apples for the transient voltage term V1 and the
estimation of the internal resistance parameter R1. To be able to separate these effects, setting
different measurement noises for the state- and weight filters might be very useful. As is known, the
elements of the state vector will change much faster than those in the weight vector. This is reflected
by the weight vector being initialized with a larger measurement noise covariance than the state filter.
87
The balance between the measurement noises was found to be critical for the proper convergence of
the filter. It should however be noted that a similar separation-of-time scales effect could be achieved
by setting the measurement noises identically for the two filters and then lowering the values in the
weight filter process noise covariance matrix accordingly. Examples of this may be seen in the results
section 4.6.2 below. Campestrini et al. applied different measurement noise covariance matrices in the
state- and weight filters of their DEKF approach without motivation.
The process noise covariance matrix plays quite a different role in the weight filter than in the state
filter as it is used to “drive” the process, for the lack of a better weight filter state equation. This was
especially evident for the estimation of cell capacity. Setting the capacity process noise covariance too
high resulted in a too fast capacity degradation as shown in the figure below. The opposite was also
applicable, setting the weight filter process noise covariance matrix too low results in no changes to
parameter values over time at all. Another phenomenon could be observed if the state filter SoC
estimation process noise covariance was set low while the weight filter capacity estimation process
noise covariance was set to high values. This would lead to estimates of both SoC and capacity
diverging. This is because the capacity estimate is overcorrected to adjust for errors in SoC, but the
algorithm still considers the SoC estimate from the state filter to be relatively accurate due to the low
SoC covariance. The result is a self-accelerating effect where the capacity estimates become
increasingly inaccurate causing the SoC estimates to diverge without the state filter “intervening”
enough to adjust the faulty SoC estimate.
Figure 57: Capacity estimation using the DEKF over the first RW phase of RW9 with the capacity process noise covariance set too high. Internal filter parameters are like in table 12 except for capacity process noise covariance=1∙10-8.
Again, the capacity process noise covariance was found to be especially important. For the internal
resistances R0 and R1, the relevant values of the process noise covariance matrix seemed to mostly
control the noise of the estimation, meaning that appropriate values of these parameters could be
found by simple visual inspection. This can be seen in the figure below where estimations of internal
resistances have been performed using identical internal filter parameters to those used to obtain the
results in figure 51 bar the values of process noise covariance for the internal resistance parameters.
88
Note how effect of varying the process noise covariances only become evident after some time has
passed. This is due to initial covariance matrices being set identically for all simulations in figure 58.
Figure 58: Convergence of estimates of internal resistance parameters for varying values of process noise covariance in the DEKF.
As is evident, a higher process noise covariance leads to a noisier estimate of the internal resistance
parameters. It is notable how the relation between the process noise covariance and noise level of the
estimation appears very similar for R0 and R1. For this reason, it seemed sensible to keep the values in
the process noise covariance matrices identical for these.
Finally, it is interesting to note that estimation of internal resistances R0 and R1 was much more
straightforward then the estimation of the cell capacity from a filter tuning perspective. The DEKF was
found to estimate these parameters with a level of robustness that was simply not possible for the cell
capacity, especially so when data from battery RW2 was used. It is thought that this might be due to
the “entanglement” effect of the SoC and the cell capacity as mentioned above. From this perspective,
investigation of a correlation between cell capacity loss and internal resistance gain due to aging would
possibly be of high interest. Theoretically, online estimation of internal resistance parameters should
also allow for more accurate short-term SoC estimation as effects of, say, changing temperatures could
be indirectly controlled for.
4.4 ENHANCED STATE VECTOR EXTENDED KALMAN FILTER The first of the suggested simpler Kalman filter SoH estimation algorithms, the ESVEKF, uses a single
state filter to estimate the cell capacity. Due to the similarity to the already analyzed EKF, a lot of
lessons learned in studying that approach could be applied here as well. The filter was found to be able
to estimate the cell capacity rapidly and accurately despite poor initializing values, as can be seen
below. The ESVEKF does not estimate the internal resistance parameters of the cell. Because of this
fact, the measured values from pulsed discharge tests were used for all results. The results in figure 58
are quite similar in terms of accuracy and overall appearance to what was observed for the DEKF in
figure 49.
89
Figure 59: Capacity estimation over entire first RW phase of battery RW9 using ESVEKF.
After the initial stabilizing period, SoC estimation is accurate and remains so for the duration of the
first RW phase:
Figure 60: SoC estimation using the ESVEKF for a poor initial guess of cell capacity.
For filter tuning of the ESVEKF, the found-to-be successful internal filter parameter values of the EKF
were the starting point. Indeed, the most fruitful results of the ESVEKF were achieved with internal
filter parameters set very similarly. The values of internal filter parameters used for results shown in
figures 59 and 60 are summarized below.
90
Table 15: Internal filter parameters of the ESVEKF for the results in figures 59 and 60.
State filter internal parameters
Starting SoC 0.999
Starting V1 (V) 0.00 V
Starting C 1.9 Ah
Covariance, SoC 1∙10-4
Covariance, V1 1∙10-6 V2
Covariance, 1/C 1∙10-4 (1/Ah)2
Process noise, SoC 1∙10-10
Process noise, V1 1∙10-4 V2
Process noise, 1/C 1∙10-11 (1/Ah)2
Measurement noise 5∙10-3 V2
The number of internal filter parameters has been reduced from the 17 of the DEKF to just 10 in the
ESVEKF, which is just three more than the EKF. Due to the single filter approach, only one measurement
noise covariance value must be assigned, making finding an appropriate value for this parameter as
intuitive as in the EKF. The algorithm was found to be quite insensitive to the starting guess of cell
capacity, converging to a capacity value close to the measured 2.10 Ah for a wide array of different
initializing values:
Figure 61: Capacity estimation using the ESVEKF converging to measured value independently of starting guess.
For comparisons sake, all internal filter parameters except the starting guess of capacity were the same
in figure 61 as in figures 59 and 60. The ESVEKF would also quickly converge even if the starting guess
of SoC also was uncertain. In the figure below the initial capacity was set to 1.9 Ah while the initial SoC
was set to 0.5. All other internal filter parameters were the same as for previously shown ESVEKF
results. Just like for the DEKF, the capacity and SoC estimates converge approximately simultaneously
and SoC estimation remains close to the Coulomb counting reference thereafter.
91
Figure 62: Convergence of SoC using the ESVEKF despite poor guesses of both SoC and capacity.
Like for the DEKF, data from the battery RW2 was also examined in the ESVEKF. Also here capacity
estimates quickly converge and the algorithm appears relatively robust to initialization. Of course,
unlike for the DEKF, initial values of internal resistance parameters and capacitance were required.
Again, the method based the MATLAB function “fminsearch”, as outlined in section 3.1.3, was applied
to identify these values. The measured ECM parameters of battery RW2 before cycling began were:
Table 16: Initial ECM parameters of battery RW2.
R0 R1 C1
0.099 Ω 0.055 Ω 1447 F
Figure 63: Rapid convergence of capacity estimates with three different initial guesses using ESVEKF and data from battery RW2
92
It is seen in figure 63 that convergence is very rapid and virtually independent of the initial guess of
capacity, which were 3.0 Ah, 2.0 Ah, and 1.0 Ah here, respectively. Like for the DEKF, capacity estimates
are initially relatively noisy, but then stabilize quite well. Again, the ESVEKF was found to be more
robust with regards to initialization than the DEKF, perhaps simply since there are fewer internal filter
parameters which to manipulate.
It is likely that periodic calibration of the ECM parameter values is necessary for the long-term use of
the ESVEKF. This is evident when the output equation of the Kalman filter is considered. Only the OCV-
term is adjusted for aging effects in the ESVEKF. The effects of this will be analyzed further in section
4.6.2. Short-term, noting again that the rate of battery aging is highly accelerated in the investigated
data compared to for most normal Li-ion battery applications, convergence to measured values was
found to be rapid and accuracy of the capacity estimation was good compared to values measured in
reference discharge tests. In addition, the filter tuning process was heavily streamlined compared to
for the DEKF due to the much smaller number of internal parameters to be considered. The more
intuitive filter tuning process might well be a significant advantage in real-life applications.
4.5 SINGLE WEIGHT DUAL EXTENDED KALMAN FILTER The SWDEKF uses the DEKF algorithm but with a weight vector consisting of only the cell capacity,
meaning that neither internal resistance- nor capacitance parameters are adjusted during filtering. The
purpose of this truncation of the weight vector was to reduce the computational strain compared to
the full DEKF whilst still achieving results of acceptable quality in terms of estimation of cell capacity
and, thus, state of health. While the number of variables being adjusted is the same as in the ESVEKF,
they are here split into two filters, which should theoretically allow for more accurate tuning of overall
estimation accuracy as well as improved computational performance. On the other hand, the
increased number of filter tuning parameters makes the tuning process more demanding and, possibly,
time consuming. As is evident from the figure below, the cell capacity estimates using this filter could
also quickly converge to the measured values despite a poor initialization. Just like for the ESVEKF, the
internal resistances and capacitances as measured in the pulsed discharge tests were used for all
SWDEKF work.
Figure 64: Convergence of capacity estimate for SWDEKF despite poor initial guess.
93
It is interesting to note the two apparent plateaus in the capacity estimation, very similar to the DEKF
results pictured in figure 49. The following internal filter parameters were used to obtain the results
above:
Table 17: Internal filter parameters for SWDEKF results in figure 64.
State filter internal parameters
Starting SoC 0.999
Starting V1 0.0 V
Covariance, SoC 1∙10-8
Covariance, V1 1∙10-4 V2
Process noise, SoC 1∙10-15
Process noise, V1 1∙10-9 V2
Measurement noise, state filter 5∙10-6 V2
Weight filter internal parameters
Starting capacity 1.9 Ah
Covariance, capacity 5∙10-6 (Ah)2
Process noise, capacity 1∙10-10 (Ah)2
Measurement noise, weight filter 1∙10-3 V2
As was the case for the DEKF, balancing the weight- and state filter internal filter parameters was
found to be critical for the overall performance of the SWDEKF. In fact, the overall behavior of the filter
was very similar to that of the DEKF. Rate of convergence to measured capacity values was quite rapid
for a wide array of starting guesses. The figure below displays the convergence behavior of SWDEKF
capacity estimates for a wide array of starting guesses of cell capacity for the start of the first RW phase
of battery RW9.
Figure 65: Convergence of capacity estimates to measured value for RW9 for different starting guesses in SWDEKF.
It can be seen when comparing figures 65 and 61 that the ESVEKF and the SWDEKF appear to converge
at approximately the same rate, around three hours into the RW phase. The spread of estimated
capacities for the different starting guesses is slightly wider for the SWDEKF. This might however be
an issue of filter tuning and could possibly be improved by enough trial-and-error testing of internal
filter parameters.
94
Again, the cycling data from battery RW2 was also used with the SWDEKF. While results were found
to be like those for the DEKF and the ESVEKF, convergence was found to be generally slower, as is
evident in the figure below.
Figure 66: Capacity estimates for battery RW2 using the SWDEKF for different starting guesses of capacity.
Unlike for the other two investigated capacity estimation algorithms, this slow convergence behavior
was not helped by increasing the initial capacity covariance. This would only have the effect of
introducing more turbulent behavior and spiking at the start of filtering, actually slowing down
convergence compared to when using a relatively low initial capacity covariance such as in figure 66.
Figure 67: Capacity estimates using SWDEKF for RW2 for high initial capacity covariance.
95
In terms of long-term capacity estimation, the same concerns as for the ESVEKF apply. The results of
short-term capacity estimation simulations where overall slightly worse for this algorithm than for
ESVEKF in terms of accuracy and convergence rate. Again, this might be due to filter tuning issues
however. The higher number of internal filter parameters means that finding appropriate values of
these was more difficult than for the ESVEKF.
4.6 COMPARISON OF ALGORITHMS AND SUMMARY
4.6.1 State of charge algorithms
The two investigated SoC estimation algorithms both performed very well in terms of accuracy to the
Coulomb counting reference. To compare the accuracy of SoC estimates quantitatively, internal filter
parameters were set up identically for the two filters, with the starting values of process- and
measurement noise covariance matrices for the AEKF being the same as the equivalent constant values
in the EKF. The filters were then used to simulate the 300 first RW steps of the first RW phase of battery
RW9, which is made up of approximately 70 000 data points. The 300 first RW steps were chosen for
this comparison as this amount of data allows for differences in accuracy between the algorithms to
be clearly distinguishable but it still early enough in the RW phase to be able to reasonably assume
that the values of ECM parameters are roughly the same as before cycling began, meaning that
Coulomb counting represents a good reference for SoC estimates. Choices of internal parameters for
the two filters are shown in the table below:
Table 18: Internal filter parameters for quantitative comparison of EKF and AEKF algorithms.
State filter internal parameters EKF AEKF
Starting SoC 0.5 0.5
Starting V1 0.0 V 0.0 V
Covariance, SoC 1∙10-2 1∙10-2
Covariance, V1 1∙10-3 V2 1∙10-3 V2
(Starting) Process noise, SoC 1∙10-10 1∙10-10
(Starting) Process noise, V1 1∙10-4 V2 1∙10-4 V2
(Starting) measurement noise 5∙10-3 V2 5∙10-3 V2
Moving estimation window size, M - 10
In addition, artificial random noise was added to measured cell voltage and current signals on the
form of (68) with standard deviations of 5∙10-3 V and 5∙10-3 A. ECM parameters were set as measured
values in pulsed- and reference discharge tests before the first RW phase of RW9 for both compared
algorithms. This setup produced the following results:
Table 19: Results of comparison of EKF and AEKF for SoC estimation
EKF AEKF
Root-mean-square SoC error4 0.0142 0.0176
Execution time (s) 5.9 7.5
Looking at table 19, the results point towards the EKF being the superior alternative, both producing a
more accurate result and being faster to execute. What is not included in table 19 however is the time
needed for filter tuning. This process was much easier for the AEKF as parameters for process- and
measurement noise covariance are updated recursively during filtering, meaning that initialization was
less sensitive. Of course, the window size needs to be set in the AEKF as well, but as discussed above
this was not a major issue and its value had relatively little influence on the accuracy of estimation
4 Compared to the Coulomb counting reference.
96
results. It may well turn out to be advantageous to apply an AEKF in situations where initialization is
difficult to rationalize or where earlier operation data is scarce. For the purposes of this work the added
accuracy and speed of the EKF was prioritized higher than robustness to initialization however,
especially considering how much emphasis was put on characterizing filter behavior and filter tuning.
4.6.2 State of health estimation algorithms
It is clear that all investigated SoH estimation algorithms were capable of estimating cell capacity
during online operation. This was the case even if initial capacity guesses were very inaccurate, which
would result in all algorithms relatively quickly converging to measured values of capacity, given that
all other ECM parameters were correctly set in accordance with measured values. The DEKF also had
the advantage of being able to estimate the internal resistance parameters of the cell, which should
increase accuracy of estimations both short- and long term due to the increased model accuracy. As
all model parameters, see table 8 for the current case of a 1 RC ECM, change slowly over time as the
battery ages, the accuracy of the ESVEKF and SWDEKF should falter over longer periods. In addition,
internal resistance is known to be a function of both temperature and current which theoretically
should allow for the DEKF to be more accurate over shorter time periods as well. This is especially the
case if conditions during measurements of model parameters are substantially different than during
operation such as the case for capacity measurements of battery RW9.
One recurring issue when discussing SoH estimation methods is the question of validation. In the case
of SoC estimation, this can be measured accurately under certain operation conditions using laboratory
equipment by Coulomb counting, at least over shorter periods of time when cell capacity can be
assumed to remain constant. In other words, a clear reference point exists for which to aim. Such a
reference method does not exist for online capacity- and/or internal resistance estimation as these
quantities can only be calculated using data taken from special operation modes, such as the pulsed-
and reference discharge tests used in this work. Two methods of validation, as outlined in section 3.2,
were used in this work, utilizing the fact that cell capacity was measured in reference discharge tests
before and after any given RW phase. Using both methods of validation, all algorithms could estimate
the capacity changes over a full, intense usage period of over 90 hours for battery RW9, as well as the
over longer time periods. It was found that differences in SoH estimation between the investigated
algorithms became more significant for longer simulation periods.
To try to quantitatively compare the algorithms, a case study consisting of the long-term capacity
estimation for battery RW2 was considered. The purpose of this test is to observe and compare the
long-term behavior of the filters in a more realistic prolonged battery monitoring situation. The reason
for using data from this battery and not RW9 was due to the more stable estimation results obtained
using RW2. The reasons for this higher stability, or robustness with regards to initialization, are
believed to be the prolonged periods of charging and lack of points of negative SoC, as calculated by
Coulomb counting, in the data. The internal filter parameters of the DEKF and SWDEKF were set
identically in terms of the shared internal filter parameters for the results below to try to make the
comparison as fair as possible. The internal filter parameters of the algorithms for this long-term test
are summarized in table 20 below. The estimated capacities of RW2 using the three algorithms for data
recorded over more than 1200 hours of RW cycling are shown in the figure below.
97
Table 20: Internal filter parameters for comparison of long-term capacity estimation performance of SoH algorithms.
Figure 68: Comparison of long-term capacity estimation of SoH algorithms.
It is of course the case that filter tuning will affect the results in figure 68. The results should be
considered merely representative of a typical situation and have not been optimized with regards to
filter tuning. It is seen that DEKF performs the best in terms of accuracy, followed by SWDEKF and then
ESVEKF. Root-mean-square errors of the SoH estimates for the algorithms are shown in table 21. The
error was defined as the difference between the measured and estimated capacities at the end of a
given RW phase. The starting capacity of RW2 was 1.98 Ah, meaning that the measured SoH at the end
of the data in figure 68 is 77 %. The test above thus represents more than a typical full Li-ion battery
lifetime in an automotive application.
State filter internal parameters DEKF SWDEKF ESVEKF
Starting SoC 0.999 0.999 0.999
Starting V1 (V) 0.0 0.0 0.0
Covariance, SoC 1∙10-5 1∙10-5 1∙10-5
Covariance, V1 (V2) 1∙10-12 1∙10-12 1∙10-12
Process noise, SoC 1∙10-10 1∙10-10 1∙10-10
Process noise, V1 (V2) 1∙10-5 1∙10-5 1∙10-5
Measurement noise, state filter (V2) 1∙10-10 1∙10-10 1∙10-10
Covariance, 1/C (1/Ah)2 - - 1∙10-14
Process noise, 1/C (1/Ah)2 - - 1∙10-16
Weight filter internal parameters
Covariance, capacity (Ah)2 1∙10-10 1∙10-10 -
Covariance, R0 (Ω2) 1∙10-12 - -
Covariance, R1 (Ω2) 1∙10-12 - -
Process noise, capacity (Ah)2 1∙10-16 1∙10-16 -
Process noise, R0 (Ω2) 1∙10-18 - -
Process noise, R1 (Ω2) 1∙10-18 - -
Measurement noise, weight filter (V2) 1∙10-10 1∙10-10 -
98
Table 21: Root-mean-square capacity estimation errors for results in figure 65 for investigated SoH estimation algorithms.
Algorithm RMS error with regards to capacity as measured in reference discharge tests
DEKF 2.2 %
SWDEKF 8.3 %
ESVEKF 10.4 %
The tests of long-term capacity estimation above were performed in the same way for all compared
algorithms. The filters were initialized with the ECM parameter values and cell capacity as measured
for RW2 in reference- and pulsed discharge tests before cycling began. These initial ECM parameter
values may be found in table 16 above. The filters were then used to simulate the entire RW phases,
that is until the next batch of cell diagnostic tests were performed. The estimated values of cell capacity
at the end of said RW phases were then used as initialization values for the simulation of the
subsequent RW phase. In addition, the initialization values of internal resistance parameters R0 and R1
were also updated in the same way as the capacity for the DEKF. The filters which do not adapt internal
resistance values during the filtering process, the ESVEKF and SWDEKF, used the initial values of these
parameters throughout testing for fairness of comparison.
As can be seen in figure 68, all algorithms perform similarly in terms of accuracy of estimation to
measured capacities at the start of the testing, especially the ESVEKF and the DEKF. The difference in
error compared to the measured values of capacity was indeed much smaller over the first four data
points, measured over more than 500 hours of simulated operation, than during the full test for all
algorithms:
Table 22: Root-mean-square error over first four measurement points of long-term capacity estimation test.
Algorithm RMS error with regards to capacity as measured in first four reference discharge tests
DEKF 1.3 %
SWDEKF 5.9 %
ESVEKF 1.9 %
For later data points, the capacity estimation accuracy of especially the algorithms not adjusting for
changes in internal resistance suffers. That said, the accuracy of the DEKF capacity estimate also
appears to decrease towards the later RW phases. While the lack of adjustment of the capacitance
parameter in any investigated algorithm likely plays a role, temperature effects might also be
important. It should however be noted that the change in average temperature during cycling over
time is much smaller for battery RW2 than it was for RW9. This is likely due to the lower average
current applied during charging of RW2 using the CC-CV mode as Joule heating (see (3)) is proportional
to the applied current squared.
At first glance it might appear odd that not adjusting for changes to internal resistance parameters
would result in too high capacity estimates for later data points, as seen for the ESVEKF and SWDEKF
in figure 68. Looking at the output equation the opposite effect seems expected. When internal
resistance parameters R0 and R1 are not adjusted to account for aging effects they will quickly be
smaller than the values one would measure for the real cell in a pulsed discharge test or estimate with
the DEKF. Disregarding changes to the capacitance parameter due to aging, the absolute values of the
terms in the output equation which are being multiplied by R0 and R1 will both be too small. For
illustrations sake, the output equation is rewritten as such by utilizing (14) and (15):
99
𝑂𝐶𝑉(𝑆𝑂𝐶) = 𝑉𝑡 + 𝑅𝑜 ∙ 𝐼 + 𝑅1 ∙ 𝐼 ∙ (1 − 𝑒
−∆𝑡𝑅1∙𝐶1) + (𝑡𝑒𝑟𝑚 𝑛𝑜𝑡 𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡 𝑜𝑛 𝐼) (69)
Given that all Kalman filter algorithms will adjust the state- and, in the case of the DEKF and SWDEKF
at least, the weight vector to make the predicted terminal voltage Vt be as close to the measured value
as possible in each time step, the terminal voltage term in (69) may be seen as fixed in this discussion.
Now consider that the situation can be divided into two cases: during discharge (current is positive)
and during charge (current is negative). For discharge with too low, unadjusted values of R0 and R1, it
is seen that the estimated OCV would need to be lower for the same cell voltage Vt to be predicted.
This is of course equivalent to a lower estimated SoC. During discharge this would mean that the cell
capacity would need to be lower to achieve a lower SoC for a given discharge current and duration.
For the other case, that of charging the cell, a similar effect should be expected. For non-adjusted
values of R0 and R1, the OCV, and thus SoC, would need to be higher for the same given terminal voltage
prediction. For charging this is equivalent to a larger positive change in SoC for the same current and
time, meaning a smaller capacity again be expected to be predicted here as well. So why is the opposite
effect observed, that is capacity estimates of the algorithms that do not adjust for changes to internal
resistance parameters becoming increasingly too large over time?
The explanation is believed to lie in the allocation of modelling error in the state process noise
covariance matrix. As mentioned in the discussion of the EKF, it is intuitive to consider the estimate of
the transient voltage V1 to be less accurate than that of the SoC and to thusly reflect this in the process
noise covariance matrix by specifying a larger value for the transient voltage covariance. This appears
to be fine for the case of a relatively accurate model with accurately specified parameters. In the case
of the ESVEKF and SWDEKF, where the internal resistance parameters are not adjusted over time, this
will however have the effect of “swallowing up” errors of the SoC- and capacity estimation as modeling
error increases. Essentially, the modelling error of the V1-term will increase much faster compared to
the OCV-term over time when changes to internal resistance parameters are not considered. Examine
the output equation rewritten yet again:
𝑉𝑡 + 𝑅0 ∙ 𝐼 = 𝑂𝐶𝑉(𝑆𝑂𝐶) − 𝑉1 (70)
For discharge the left side of (70) will be too small and during charging it will be too large because of
the error in R0. Due to the process noise covariance being set to a larger value for V1 than for SoC, it
will be adjusted more to make up the difference to Vt. What is observed then is a smaller change in
OCV for a given current over a certain time, leading to a larger cell capacity being predicted, and thus
explaining the observed effect. One could consider it a misallocation of the error due to the balance of
modeling error being shifted unevenly as the battery ages. This might also explain why the same effect
is seen to a smaller extent in the DEKF as well as it still is a question of V1 becoming increasingly
inaccurate over time for this algorithm, but here due to an inaccurate value of the capacitance
parameter. Incorporating the capacitance into the weight vector of the DEKF might improve long-term
capacity estimates for this reason.
Overcoming this type of issue within the context of the ESVEKF and SWDEKF algorithms appears
difficult, even though the effects could likely be made less impactful through careful filter tuning. The
issue of the shifting of modeling error seems to be an inherent problem of all Kalman filtering
algorithms. Perhaps some adaptive solution for online estimation of process noise could be the
solution as the problem appears to stem from the assignment of constant process noise covariance
matrices. It might also be worth considering the approaches suggested by Zou et al. and Hu et al. which
incorporate a difference in time scales between the state and weight filters.
100
While the DEKF displayed the most accurate results both long-term and short-term, it was also the
most computationally taxing algorithm. A comparison of the three SoH estimation methods in terms
of execution time is shown in the table below.
Table 23: Execution times of SoH estimation algorithms for simulations of 300 RW cycles.
Algorithm Execution time (s) Percentage improvement
DEKF 10.7 0 %
ESVEKF 6.5 39 %
SWDEKF 8.5 21 %
The algorithm execution time was recorded using the “timeit” MATLAB function for a simulation over
the 300 first RW steps of the first RW phase of battery RW9. The displayed values were the average
execution time of one hundred such simulations. There is clearly a substantial difference in algorithm
computational performance. It is interesting to note that dividing the burden into two filters, such as
in the SWDEKF, instead of combining into one larger filter, such as in the ESVEKF, does not appear to
be faster, despite the reduced matrix dimensions as discussed in section 3.3.5. For the application
investigated in this study, it appears that the calculation of the recursive derivatives in the SWDEKF
outweighs the reduced strain of matrix operations. This seems to be confirmed by the relatively small
difference in execution time between the SWDEKF and the DEKF as well. It is thought that the
computational advantages of the dual filter approach would be more evident for larger state- and/or
weight vectors. Some general observations when studying the SoH estimation algorithms will now be
discussed.
One recurring phenomena discovered when trying to estimate the cell capacity in an online mode is
the need for relatively quick convergence. It was observed that when intentionally providing the filter
with inaccurate initial guesses of cell capacity, the filters would often diverge unless relatively quick
convergence to measured values was ensured by adjusting the initial capacity covariance parameter.
This effect was found to be much more prominent when using the data from battery RW9 than when
using the CC-CV mode charging data from RW2. The reason for this need for quick convergence is likely
the accumulative nature of Coulomb counting, as it is featured in the state equation of all filters. If the
initial guess of the cell capacity is lower than the measured value, the filter SoC estimate will get further
and further away from the Coulomb counting reference SoC values calculated using the measured
capacity during a given period of constant current charge or discharge such as during the RW phases
of battery RW9. For the Kalman filter SoC estimate to approach the reference values the capacity
estimate would now need to assume a higher value than what was measured. This can be explained
mathematically by noting that one divided by the capacity will be directly proportional to the slope of
the SoC-time curve when current is held constant, meaning that the sum of all currents over said period
is also a constant:
𝑆𝑂𝐶𝑘+1 = 𝑎 + 𝑏 ∙ 𝑡 = 𝑆𝑂𝐶𝑘 +∑ 𝐼𝑘 ∙ ∆𝑡𝑘𝑘
𝐶= 𝑆𝑂𝐶𝑘 + (
𝐼𝑠𝑢𝑚
𝐶) ∙ 𝑡𝑠𝑢𝑚 (71)
This overcompensation of the capacity estimate might cause serious convergence issues and should
preferably be avoided by ensuring either that the initial capacity estimate is reasonably accurate or by
careful tuning of the initial capacity covariance. If filter tuning is reasonable, capacity estimates for
cases of a faulty initial guess of the capacity are therefore expected to oscillate to some extent before
converging. This principle is only strictly applicable for cases of periods of constant current however.
In the data used in this work, the applied current was very dynamic, meaning that this effect should
101
be reduced. This is because the SoC estimated with a faulty capacity might “accidentally” coincide with
the correct SoC value when the applied current changes. This principle is shown in the figure below.
Figure 69: Principle of "accidental" convergence of SoC estimation for faulty concurrent estimation of capacity. The red line represents an SoC estimate with a low capacity, the green line high capacity, and the blue line is the in-between case.
In figure 69 the blue line represents the “true” SoC, which could for instance be represented by
Coulomb counting using a correct capacity. The red line represents a concurrent estimate of SoC and
cell capacity where the capacity is too low and the green line the same situation but with the capacity
being estimated as a too high value. During the initial 15 minutes of constant current charging, the SoC
estimates are diverging due to the variations in capacity. As the current changes to a discharging mode,
as after 15 minutes in the figure above, the SoC estimates will however tend to converge again. Note
that time and SoC in figure 69 are only there for the sake of illustration of the discussed principle and
do not represent realistic values.
This overcompensation phenomenon is also likely part of the reason why estimation of internal
resistance parameters appeared more robust than estimating cell capacity as the cell voltage response
to changes of these parameters is immediate, as seen in figure 27. It also means that it is more difficult
to estimate the cell capacity under conditions of rapidly varying current. This is because changes in cell
capacity become less noticeable per the principle shown in figure 69. In short, there is not enough time
for the inaccurate slope of the SoC-time curve to cause the SoC estimate to drift substantially from the
correct value, given that the starting SoC is reasonably accurate. And of course, it is only errors in SoC,
leading to errors in OCV, which are processed by the Kalman filter algorithm to update estimates of
state and system weights. This would also partly explain the observed more stable capacity estimation
behavior when using data from battery RW2. The relatively long periods of charging in a CC-CV mode
provides the algorithm enough time to accurately adapt the capacity estimation.
Finding the suitable internal parameters was cumbersome for the SWDEKF. The algorithm lacks the
simplicity of the ESVEKF and the tunability of the full DEKF. As modelling errors inevitably increase
during cycling in the RW phases, all adjustments must be included in the capacity estimation. This
meant that estimates tended to either be very noisy, perhaps due to the algorithm including short
term variations in internal resistance into the capacity estimate, or to change too quickly over time
unless choices internal parameters were considered very carefully. It appeared that this issue was of
102
less concern for the ESVEKF which produced more stable results. This is believed to be due to the
capacity- and SoC estimates sharing the same covariance matrix due to the joint estimation approach.
While initializing state covariance matrices were always chosen to be diagonal this property would
quickly be lost during the process of filtering due to the state covariance measurement update
equation (28). The covariance of the SoC- and capacity estimate, meaning the value of positions (2,3)
and (3,2) in the state covariance matrix of the ESVEKF, might be a very important parameter for
accurate and robust capacity estimation. This is of course missing in both dual filter approaches. The
loss of information on relevant cross-correlations between state- and weight estimations in the duel
filter approaches was also speculated to possibly decrease overall algorithm accuracy by Plett [150].
Also, likely due to the expected increase of process noise during operation as modelling errors become
more severe, which is of course very difficult to predict quantitatively, it was found to be much easier
to achieve reasonably accurate short-term capacity estimates for all investigated SoH estimation
algorithms. More specifically, towards the beginning of a given RW phase. This was especially evident
when using the algorithms that just adjusted the cell capacity. Quick convergence of the capacity
estimate could be ensured, as discussed above, quite handily, but all the capacity estimates of
suggested algorithms tended to drift over longer time periods unless the filter parameters were tuned
in hindsight. Again, the ESVEKF seemed to perform well in this regard and would converge very quickly
to measured values in these types of scenarios were all ECM parameters except the cell capacity were
set to the latest measured values.
Looking at filter tuning more generally, it is easy to confirm the conclusion of Campestrini et al. that it
is essentially important for estimation results of Kalman filtering approaches [130]. While mostly
algorithms for SoC estimation were investigated in that work, it has been seen here that filter tuning
is as, if not even more, important for SoH estimation as well. This was most striking for the dual filter
approaches, the DEKF and the SWDEKF. Not only do both filters involve many internal filter
parameters, they also require the “balancing” of the two filters. Because the state- and weight filters
in the DEKF and SWDEKF share the same output equation, it becomes very easy to overcompensate
for errors in the state filter by excessive adjusting of the weight filter. In this work, the balancing of the
two filters to achieve realistic results was performed essentially by combining known parameter values
from literature and trial-and-error, utilizing visual inspection. This appears inefficient and the filter
tuning should be studied further to find a way of streamlining the process. As the ESVEKF only utilized
a single filter it did not suffer from these issues of difficult filter tuning to the same extent. As it stands,
this is a major advantage of the ESVEKF, especially considering that results were comparable to those
obtained with DEKF and SWDEKF, at least for short-term estimation.
Looking at data from battery RW2, where charging was done in a CC-CV mode, is very interesting.
While this mode of operation certainly can be considered less realistic than what was used for cycling
of battery RW9, it does allow for control of capacity estimation results through the points of 100 %
SoC. Firstly, it appears that results for all filters were more stable for RW2. It is suspected that this is
at least partly due to the lack of negative SoCs, as discussed briefly in section 3.2, put perhaps also
because the relatively long charging periods allow for capacity estimates to stabilize, as mentioned in
the discussion of (67). The issues of negative SoCs for RW9 could possibly be resolved by integrating a
temperature- and/or current dependence for the cell capacity. As, again, the data from RW9 attempts
to resemble a realistic scenario, it appears that establishing these types of functional dependencies is
of high importance, not only for the accuracy, but also for the stability of monitoring algorithms. The
higher stability of results is evident in the apparent ability of capacity estimates to converge slowly to
the right values, thus seeming to at least partly avoid the issues of having to ensure quick convergence
as discussed above.
103
5 CONCLUSIONS AND FUTURE WORK
Two methods of SoC estimation and three methods of SoH estimation were examined in this work. All
were based on Kalman filtering approaches. Both investigated SoC estimation algorithms, the
extended Kalman filter and the adaptive extended Kalman filter, performed similarly in terms of
accuracy to a Coulomb counting reference. Both were also proven to be able to overcome most of the
commonly cited shortcomings of Coulomb counting. It was however found that the extended Kalman
filter could be made to achieve both a slightly higher accuracy and faster computational performance
if internal filter parameters were adjusted appropriately. As the process of finding suitable values for
these internal filter parameters was rather painless in the case of the EKF this algorithm was found to
be more promising than the AEKF for the application of Li-ion battery SoC estimation. It was however
noted that the accuracy of SoC estimation will quickly degrade no matter what regression algorithm
was used if the cell capacity was not adjusted to account for effects of battery aging.
The SoH algorithms were all able to accurately estimate the cell capacity over short time scales. The
ESVEKF was found to be the easiest to use from a filter tuning perspective due to the lower number of
internal filter parameters and was also found to be slightly faster than the dual filter based SWDEKF.
In addition, the ESVEKF capacity estimate could be made to converge the fastest to measured
capacities of all algorithms. In terms of long-term, continuous capacity estimation the DEKF algorithm
performed the best of all investigated algorithms by far. It is believed that this is due to this algorithm
also adjusting for changes to internal resistance parameters as the battery ages. The impact of filter
tuning on estimation of especially cell capacity was found to be especially substantial, much more so
than for online estimation of internal resistance, another common state of health indicator. The
difficulty of achieving robust capacity estimation is believed to be due to the entangled effects of the
state of charge and the cell capacity on the cell voltage but this needs to be investigated further. Time-
scale separation of filters or application of a separate output equation for the weight filter appears
promising routes to explore in this regard. The relative ease of estimating internal resistance
parameters also promotes the idea of developing some more general idea of state of health and
thereby finding some correlation between internal resistance gain and capacity fade.
As this work was only concerned with simulations, real online applications must be explored further.
This includes integration of algorithms into some sort of BMS with the goal of online capacity- and/or
internal resistance monitoring. As these parameters are normally not measurable during actual battery
operation, incorporating such an algorithm would allow not only for the estimation of battery SoH, but
also allows for the automated collection of monitoring data. Currently, the largest obstacle that still
must be overcome in developing a more holistic battery aging model capable of predicting the rate of
battery aging as a function of operation conditions, such as temperature and usage profile, is the need
for huge amounts of laboratory data. This data needs to be obtained from carefully organized vast test
matrices, covering all possible operation scenarios. Due to the slow nature of battery aging this data
acquisition is very time consuming. Currently, the answer to this has been to perform so-called
accelerated aging tests, which might instead fail to approximate real battery operation faithfully
enough. Online, continuous, recording of capacity degradation or internal resistance increase along
with the traditional parameters of current, voltage and temperature would allow for the establishing
a database. This database could then be used to develop an aging model, perhaps using a methodology
based on machine learning principles.
Another important factor to consider is the transition of monitoring algorithms from the cell scale to
the battery pack scale. Reports on the aging behavior of entire packs or modules of Li-ion cells are rare
and lots of work remains to be done within this area [57].
104
6 REFERENCES
1. Price, A., Electrical energy storage—a review of technology options. Proceedings of the Institution of Civil Engineers - Civil Engineering, 2005. 158(6): p. 52-58.
2. Dunn, B., H. Kamath, and J.-M. Tarascon, Electrical energy storage for the grid: a battery of choices. Science, 2011. 334(6058): p. 928-935.
3. Stan, A.-I., et al. Lithium ion battery chemistries from renewable energy storage to automotive and back-up power applications—An overview. in Optimization of Electrical and Electronic Equipment (OPTIM), 2014 International Conference on. 2014. IEEE.
4. Diouf, B. and R. Pode, Potential of lithium-ion batteries in renewable energy. Renewable Energy, 2015. 76: p. 375-380.
5. Lawder, M.T., et al., Battery energy storage system (BESS) and battery management system (BMS) for grid-scale applications. Proceedings of the IEEE, 2014. 102(6): p. 1014-1030.
6. Joffres, B., et al., Thermochemical Conversion of Lignin for Fuels and Chemicals: A Review. Oil Gas Sci. Technol. – Rev. IFP Energies nouvelles, 2013. 68(4): p. 753-763.
7. Berecibar, M., et al., Critical review of state of health estimation methods of Li-ion batteries for real applications. Renewable and Sustainable Energy Reviews, 2016. 56: p. 572-587.
8. Palacín, M. and A. de Guibert, Why do batteries fail? Science, 2016. 351(6273): p. 1253292. 9. Bole, B., C.S. Kulkarni, and M. Daigle. Adaptation of an electrochemistry-based li-ion battery
model to account for deterioration observed under randomized use. in Proceedings of Annual Conference of the Prognostics and Health Management Society, Fort Worth, TX, USA. 2014.
10. Blomgren, G.E., The Development and Future of Lithium Ion Batteries. Journal of The Electrochemical Society, 2017. 164(1): p. A5019-A5025.
11. Yoshino, A., The Birth of the Lithium-Ion Battery. Angewandte Chemie International Edition, 2012. 51(24): p. 5798-5800.
12. Yoshio, M., Lithium-Ion Batteries. Lithium-Ion Batteries : Science and Technologies, ed. R.J. Brodd and A. Kozawa. 2010, Dordrecht: Dordrecht : Springer.
13. Tsuyoshi, S., U. Yoshio, and N. Petr, Memory effect in a lithium-ion battery. Nature Materials, 2013. 12(6): p. 569.
14. Phone fire faults. New Scientist, 2017. 233(3110): p. 7. 15. Brodd, R.J., Synopsis of the Lithium-Ion Battery Markets, in Lithium-Ion Batteries: Science and
Technologies, M. Yoshio, R.J. Brodd, and A. Kozawa, Editors. 2009, Springer New York: New York, NY. p. 1-7.
16. Ciez, R.E. and J.F. Whitacre, The cost of lithium is unlikely to upend the price of Li-ion storage systems. Journal of Power Sources, 2016. 320: p. 310-313.
17. Van Noorden, R., A BETTER BATTERY. Nature, 2014. 507(7490): p. 26-28. 18. Pillot, C., Battery market development for consumer electronics, automotive, and industrial:
materials requirements and trends. The 11th China International Battery Fair, ShenZhen, China, 2014.
19. Gerssen-Gondelach, S.J. and A.P.C. Faaij, Performance of batteries for electric vehicles on short and longer term. Journal of Power Sources, 2012. 212: p. 111-129.
20. Klass, V., et al., Battery Health Estimation in Electric Vehicles. 2015, Stockholm: KTH Royal Institute of Technology: Stockholm.
21. Doughty, D.H., et al., Batteries for large-scale stationary electrical energy storage. The Electrochemical Society Interface, 2010. 19(3): p. 49-53.
22. Nykvist, B. and M. Nilsson, Rapidly falling costs of battery packs for electric vehicles. Nature Clim. Change, 2015. 5(4): p. 329-332.
23. Geth, F., et al., An overview of large-scale stationary electricity storage plants in Europe: Current status and new developments. Renewable and Sustainable Energy Reviews, 2015. 52: p. 1212-1227.
105
24. Luo, X., et al., Overview of current development in electrical energy storage technologies and the application potential in power system operation. Applied Energy, 2015. 137: p. 511-536.
25. Ahmadi, L., et al., Energy efficiency of Li-ion battery packs re-used in stationary power applications. Sustainable Energy Technologies and Assessments, 2014. 8: p. 9-17.
26. Swierczynski, M., et al., The Second Life Ageing of the NMC/C Electric Vehicle Retired Li-Ion Batteries in the Stationary Applications. ECS Transactions, 2016. 74(1): p. 55-62.
27. Heymans, C., et al., Economic analysis of second use electric vehicle batteries for residential energy storage and load-levelling. Energy Policy, 2014. 71: p. 22-30.
28. Tarascon, J.-M., et al., Advanced Li-ion, in Electrochemical Energy Storage. 2015, John Wiley & Sons, Inc. p. 11-15.
29. Jokar, A., et al., Review of simplified Pseudo-two-Dimensional models of lithium-ion batteries. Journal of Power Sources, 2016. 327: p. 44-55.
30. Goodenough, J.B., General Concepts, in Lithium Ion Batteries. 2007, Wiley-VCH Verlag GmbH. p. 1-25.
31. Yoshino, A., 1 - Development of the Lithium-Ion Battery and Recent Technological Trends A2 - Pistoia, Gianfranco, in Lithium-Ion Batteries. 2014, Elsevier: Amsterdam. p. 1-20.
32. Schipper, F. and D. Aurbach, A brief review: Past, present and future of lithium ion batteries. Russian Journal of Electrochemistry, 2016. 52(12): p. 1095-1121.
33. Broussely, M., et al., Main aging mechanisms in Li ion batteries. Journal of Power Sources, 2005. 146(1): p. 90-96.
34. Julien, C., Design Considerations for Lithium Batteries, in Materials for Lithium-Ion Batteries, C. Julien and Z. Stoynov, Editors. 2000, Springer Netherlands: Dordrecht. p. 1-20.
35. Abe, K., Nonaqueous Electrolytes and Advances in Additives, in Electrolytes for Lithium and Lithium-Ion Batteries, T.R. Jow, et al., Editors. 2014, Springer New York: New York, NY. p. 167-207.
36. Miller, B.P., Automotive Lithium-Ion Batteries. Johnson Matthey Technology Review, 2015. 59(1): p. 4-13.
37. Goodenough, J.B. and K.-S. Park, The Li-ion rechargeable battery: a perspective. Journal of the American Chemical Society, 2013. 135(4): p. 1167.
38. Thackeray, M., Lithium-ion batteries: An unexpected conductor. Nat Mater, 2002. 1(2): p. 81-82.
39. Julien, C., et al., Comparative Issues of Cathode Materials for Li-Ion Batteries. Inorganics, 2014. 2(1): p. 132.
40. Nitta, N., et al., Li-ion battery materials: present and future. Materials Today, 2015. 18(5): p. 252-264.
41. Vetter, J., et al., Ageing mechanisms in lithium-ion batteries. Journal of Power Sources, 2005. 147(1): p. 269-281.
42. Thackeray, M., et al., Electrochemical extraction of lithium from LiMn2O4. Materials Research Bulletin, 1984. 19(2): p. 179-187.
43. Padhi, A.K., K. Nanjundaswamy, and J.B. Goodenough, Phospho‐olivines as positive‐electrode materials for rechargeable lithium batteries. Journal of the electrochemical society, 1997. 144(4): p. 1188-1194.
44. Zhang, W.-J., Structure and performance of LiFePO4 cathode materials: A review. Journal of Power Sources, 2011. 196(6): p. 2962-2970.
45. Xu, T., et al., Lithium-ion batteries for stationary energy storage. JOM, 2010. 62(9): p. 24-30. 46. Waldmann, T., et al., Temperature dependent ageing mechanisms in Lithium-ion batteries - A
Post-Mortem study. Journal of Power Sources, 2014. 262: p. 129. 47. Liu, W., et al., Nickel‐Rich Layered Lithium Transition‐Metal Oxide for High‐Energy Lithium‐Ion
Batteries. Angewandte Chemie International Edition, 2015. 54(15): p. 4440-4457. 48. Swierczynski, M., et al., Selection and Performance-Degradation Modeling of LiMO2/Li4Ti5O12
and LiFePO4/C Battery Cells as Suitable Energy Storage Systems for Grid Integration With Wind
106
Power Plants: An Example for the Primary Frequency Regulation Service. Sustainable Energy, IEEE Transactions on, 2014. 5(1): p. 90-101.
49. Yoshio, M. and H. Noguchi, A Review of Positive Electrode Materials for Lithium-Ion Batteries, in Lithium-Ion Batteries: Science and Technologies, M. Yoshio, R.J. Brodd, and A. Kozawa, Editors. 2009, Springer New York: New York, NY. p. 9-48.
50. Chanson, C., Safety of lithium-ion batteries. The European Association for Advanced Rechargeable Batteries, 2013.
51. Kulova, T., New electrode materials for lithium-ion batteries (Review). Russian Journal of Electrochemistry, 2013. 49(1): p. 1-25.
52. Sandhya, C.P., B. John, and C. Gouri, Lithium titanate as anode material for lithium-ion cells: a review. Ionics, 2014. 20(5): p. 601-620.
53. Chen, Z., et al., Titanium‐Based Anode Materials for Safe Lithium‐Ion Batteries. Advanced Functional Materials, 2013. 23(8): p. 959-969.
54. Monconduit, L., et al., Negative Electrodes, in Electrodes for Li-Ion Batteries. 2015, John Wiley & Sons, Inc. p. 1-28.
55. Yan-Bing, H., et al., Gassing in Li4Ti5O12-based batteries and its remedy. Scientific Reports, 2012. 2.
56. Han, X., et al., A comparative study of commercial lithium ion battery cycle life in electrical vehicle: Aging mechanism identification. Journal of Power Sources, 2014. 251: p. 38-54.
57. Demirkan, M., L. Trahey, and T. Karabacak, Cycling performance of density modulated multilayer silicon thin film anodes in Li-ion batteries. Journal of Power Sources, 2015. 273: p. 52-61.
58. Zhang, S.S., A review on electrolyte additives for lithium-ion batteries. Journal of Power Sources, 2006. 162(2): p. 1379-1394.
59. Pop, V., Battery Management Systems. Philips Research, 9, ed. H.J. Bergveld and D. Danilov. 2008, Dordrecht: Dordrecht : Springer.
60. Vezzini, A., 15 - Lithium-Ion Battery Management A2 - Pistoia, Gianfranco, in Lithium-Ion Batteries. 2014, Elsevier: Amsterdam. p. 345-360.
61. Lu, L., et al., A review on the key issues for lithium-ion battery management in electric vehicles. Journal of Power Sources, 2013. 226: p. 272-288.
62. Ungurean, L., et al., Battery state of health estimation: a structured review of models, methods and commercial devices. 2017. p. 151-181.
63. Campestrini, C., et al., Ageing of lithium-ion battery modules with dissipative balancing compared with single-cell ageing. Journal of Energy Storage, 2016. 6: p. 142-152.
64. Rezvanizaniani, S.M., et al., Review and recent advances in battery health monitoring and prognostics technologies for electric vehicle (EV) safety and mobility. Journal of Power Sources, 2014. 256: p. 110-124.
65. Waag, W., C. Fleischer, and D.U. Sauer, Critical review of the methods for monitoring of lithium-ion batteries in electric and hybrid vehicles. Journal of Power Sources, 2014. 258: p. 321-339.
66. Friesen, A., et al., Influence of temperature on the aging behavior of 18650-type lithium ion cells: A comprehensive approach combining electrochemical characterization and post-mortem analysis. Journal of Power Sources, 2017. 342: p. 88-97.
67. Wu, C., et al., A review on fault mechanism and diagnosis approach for Li-ion batteries.(Report). Journal of Nanomaterials, 2015.
68. Victor, A. and F. Jeffrey, Lithium Ion Battery Anode Aging Mechanisms. Materials, 2013. 6(4): p. 1310-1325.
69. Birkl, C.R., et al., Degradation diagnostics for lithium ion cells. Journal of Power Sources, 2017. 341: p. 373-386.
70. Barré, A., et al., A review on lithium-ion battery ageing mechanisms and estimations for automotive applications. Journal of Power Sources, 2013. 241: p. 680-689.
107
71. Tebbe, J.L., T.F. Fuerst, and C.B. Musgrave, Mechanism of hydrofluoric acid formation in ethylene carbonate electrolytes with fluorine salt additives. Journal of Power Sources, 2015. 297: p. 427-435.
72. Lux, S.F., et al., The mechanism of HF formation in LiPF6 based organic carbonate electrolytes. Electrochemistry Communications, 2012. 14(1): p. 47-50.
73. Kanevskii, L. and V. Dubasova, Degradation of lithium-ion batteries and how to fight it: A review. Russian Journal of Electrochemistry, 2005. 41(1): p. 1-16.
74. Schultz, C., et al., Separation and Quantification of Organic Electrolyte Components in Lithium-Ion Batteries via a Developed HPLC Method. Journal of the Electrochemical Society, 2015. 162(4): p. A629-A634.
75. Deshpande, R.D. and D.M. Bernardi, Modeling Solid-Electrolyte Interphase (SEI) Fracture: Coupled Mechanical/Chemical Degradation of the Lithium Ion Battery. Journal of The Electrochemical Society, 2017. 164(2): p. A461-A474.
76. Sheldon, B.W. and A. Tokranov, Internal stress due to solvent co-intercalation in graphite electrodes for Li ion batteries. Extreme Mechanics Letters, 2016. 9, Part 3: p. 379-385.
77. Deshpande, R., et al., Battery Cycle Life Prediction with Coupled Chemical Degradation and Fatigue Mechanics. Journal of the Electrochemical Society, 2012. 159(10): p. A1730-A1738.
78. Li, Z., et al., A review of lithium deposition in lithium-ion and lithium metal secondary batteries. Journal of Power Sources, 2014. 254: p. 168-182.
79. Smart, M.C. and B.V. Ratnakumar, Effects of Electrolyte Composition on Lithium Plating in Lithium-Ion Cells. Journal of The Electrochemical Society, 2011. 158(4): p. A379.
80. Petzl, M., M. Kasper, and M.A. Danzer, Lithium plating in a commercial lithium-ion battery – A low-temperature aging study. Journal of Power Sources, 2015. 275: p. 799-807.
81. Jaguemont, J., L. Boulon, and Y. Dubé, A comprehensive review of lithium-ion batteries used in hybrid and electric vehicles at cold temperatures. Applied Energy, 2016. 164: p. 99-114.
82. Farmann, A., et al., Critical review of on-board capacity estimation techniques for lithium-ion batteries in electric and hybrid electric vehicles. Journal of Power Sources, 2015. 281: p. 114-130.
83. Belharouak, I., G.M. Koenig Jr, and K. Amine, Electrochemistry and safety of Li4Ti5O12 and graphite anodes paired with LiMn2O4 for hybrid electric vehicle Li-ion battery applications. Journal of Power Sources, 2011. 196(23): p. 10344-10350.
84. Banerjee, A., et al., Improving Stability of Li-Ion Batteries by Means of Transition Metal Ions Trapping Separators. Journal of The Electrochemical Society, 2016. 163(6): p. A1083-A1094.
85. Kim, Y., Encapsulation of LiNi0.5Co0.2Mn0.3O2 with a thin inorganic electrolyte film to reduce gas evolution in the application of lithium ion batteries. Physical Chemistry Chemical Physics, 2013. 15(17): p. 6400-6405.
86. Ecker, M., et al., Development of a lifetime prediction model for lithium-ion batteries based on extended accelerated aging test data. Journal of Power Sources, 2012. 215: p. 248-257.
87. Omar, N., et al., Lithium iron phosphate based battery – Assessment of the aging parameters and development of cycle life model. Applied Energy, 2014. 113: p. 1575-1585.
88. Schmalstieg, J., et al., A holistic aging model for Li(NiMnCo)O2 based 18650 lithium-ion batteries. Journal of Power Sources, 2014. 257: p. 325-334.
89. Thomas, E.V., et al., Statistical methodology for predicting the life of lithium-ion cells via accelerated degradation testing. Journal of Power Sources, 2008. 184(1): p. 312-317.
90. Bandhauer, T.M., S. Garimella, and T.F. Fuller, A Critical Review of Thermal Issues in Lithium-Ion Batteries. Journal of The Electrochemical Society, 2011. 158(3): p. R1.
91. Paul, S., et al., Analysis of ageing inhomogeneities in lithium-ion battery systems. Journal of Power Sources, 2013. 239: p. 642-650.
92. Su, L., et al., Path dependence of lithium ion cells aging under storage conditions. Journal of Power Sources, 2016. 315: p. 35-46.
108
93. Sarasketa-Zabala, E., et al., Calendar ageing analysis of a LiFePO4/graphite cell with dynamic model validations: Towards realistic lifetime predictions. Journal of Power Sources, 2014. 272: p. 45-57.
94. Wang, J., et al., Degradation of lithium ion batteries employing graphite negatives and nickel–cobalt–manganese oxide + spinel manganese oxide positives: Part 1, aging mechanisms and life estimation. Journal of Power Sources, 2014. 269: p. 937-948.
95. Bloom, I., et al., An accelerated calendar and cycle life study of Li-ion cells. J. Power Sources, 2001. 101(2).
96. Beltran, H., et al. Lithium ion batteries ageing analysis when used in a PV power plant. in 2012 IEEE International Symposium on Industrial Electronics. 2012.
97. Wang, J., et al., Cycle-life model for graphite-LiFePO4 cells. Journal of Power Sources, 2011. 196(8): p. 3942-3948.
98. Eddahech, A., O. Briat, and J.-M. Vinassa, Performance comparison of four lithium–ion battery technologies under calendar aging. Energy, 2015. 84: p. 542-550.
99. Belt, J.R., et al., A capacity and power fade study of Li-ion cells during life cycle testing. Journal of Power Sources, 2003. 123(2): p. 241-246.
100. Groot, J., State-of-Health Estimation of Li-ion Batteries: Ageing Models. 2014. 101. Sarikurt, T., et al., An analytical battery state of health estimation method. 2014. p. 1605-1609. 102. Nuhic, A., et al., Health diagnosis and remaining useful life prognostics of lithium-ion batteries
using data-driven methods. Journal of Power Sources, 2013. 239: p. 680-688. 103. Sarasketa-Zabala, E., et al., Realistic lifetime prediction approach for Li-ion batteries. Applied
Energy, 2016. 162: p. 839-852. 104. Nejad, S., D.T. Gladwin, and D.A. Stone, A systematic review of lumped-parameter equivalent
circuit models for real-time estimation of lithium-ion battery states. Journal of Power Sources, 2016. 316: p. 183-196.
105. Jongerden, M.R. and B.R. Haverkort, Which battery model to use? IET software, 2009. 3(6): p. 445-457.
106. Kaj, I. and V. Konané, Modeling battery cells under discharge using kinetic and stochastic battery models. Applied Mathematical Modelling, 2016. 40(17–18): p. 7901-7915.
107. Pinson, M.B. and M.Z. Bazant, Theory of SEI Formation in Rechargeable Batteries: Capacity Fade, Accelerated Aging and Lifetime Prediction. Journal of the Electrochemical Society, 2013. 160(2): p. A243-A250.
108. Santhanagopalan, S., et al., Review of models for predicting the cycling performance of lithium ion batteries. Journal of Power Sources, 2006. 156(2): p. 620-628.
109. Li, Z., et al., On state-of-charge determination for lithium-ion batteries. Journal of Power Sources, 2017. 348: p. 281-301.
110. Hu, X., S. Li, and H. Peng, A comparative study of equivalent circuit models for Li-ion batteries. Journal of Power Sources, 2012. 198: p. 359-367.
111. Lavigne, L., et al., Lithium-ion Open Circuit Voltage (OCV) curve modelling and its ageing adjustment. Journal of Power Sources, 2016. 324: p. 694-703.
112. Seaman, A., T.-S. Dao, and J. McPhee, A survey of mathematics-based equivalent-circuit and electrochemical battery models for hybrid and electric vehicle simulation. Journal of Power Sources, 2014. 256: p. 410-423.
113. Liu, G., et al. Thermal Modeling of a LiFePO 4/Graphite Battery and Research on the Influence of Battery Temperature Rise on EV Driving Range Estimation. in Vehicle Power and Propulsion Conference (VPPC), 2013 IEEE. 2013. IEEE.
114. Mishra, P.P., et al., How Does Model Reduction Affect Lithium-Ion Battery State of Charge Estimation Errors? Theory and Experiments. Journal of The Electrochemical Society, 2017. 164(2): p. A237-A251.
115. Smith, K.A., Electrochemical modeling, estimation and control of lithium ion batteries. 2006, The Pennsylvania State University.
109
116. Plett, G.L., Extended Kalman filtering for battery management systems of LiPB-based HEV battery packs: Part 1. Background. Journal of Power Sources, 2004. 134(2): p. 252-261.
117. Sun, F., et al., Adaptive unscented Kalman filtering for state of charge estimation of a lithium-ion battery for electric vehicles. Energy, 2011. 36(5): p. 3531-3540.
118. Kim, I.-S., The novel state of charge estimation method for lithium battery using sliding mode observer. Journal of Power Sources, 2006. 163(1): p. 584-590.
119. Hu, X., F. Sun, and Y. Zou, Estimation of state of charge of a lithium-ion battery pack for electric vehicles using an adaptive Luenberger observer. Energies, 2010. 3(9): p. 1586-1603.
120. Zhang, C., et al., Identification of dynamic model parameters for lithium-ion batteries used in hybrid electric vehicles. 2010.
121. Popp, H., M. Einhorn, and F.V. Conte. Capacity Decrease vs. Impedance Increase of Lithium Batteries. A comparative study. in Hybrid and Fuel Cell Electric Vehicle Symposium & Exhibition EVS26, Los Angeles. 2012.
122. Goebel, K., et al., Prognostics in Battery Health Management. Instrumentation & Measurement Magazine, IEEE, 2008. 11(4).
123. Li, J., et al., Discharge capacity estimation for Li-ion batteries based on particle filter under multi-operating conditions. Energy, 2015. 86: p. 638-648.
124. Michel, P.-H. and V. Heiries. An adaptive sigma point kalman filter hybridized by support vector machine algorithm for battery SoC and SoH estimation. in Vehicular Technology Conference (VTC Spring), 2015 IEEE 81st. 2015. IEEE.
125. Plett, G.L., Extended Kalman filtering for battery management systems of LiPB-based HEV battery packs: Part 3. State and parameter estimation. Journal of Power sources, 2004. 134(2): p. 277-292.
126. Il-Song, K., A Technique for Estimating the State of Health of Lithium Batteries Through a Dual-Sliding-Mode Observer. Power Electronics, IEEE Transactions on, 2010. 25(4): p. 1013-1022.
127. Remmlinger, J., et al., State-of-health monitoring of lithium-ion batteries in electric vehicles by on-board internal resistance estimation. Journal of Power Sources, 2011. 196(12): p. 5357-5363.
128. Zou, Y., et al., Combined State of Charge and State of Health estimation over lithium-ion battery cell cycle lifespan for electric vehicles. Journal of Power Sources, 2015. 273: p. 793-803.
129. Hu, C., B.D. Youn, and J. Chung, A multiscale framework with extended Kalman filter for lithium-ion battery SOC and capacity estimation. Applied Energy, 2012. 92: p. 694-704.
130. Campestrini, C., et al., A comparative study and review of different Kalman filters by applying an enhanced validation method. Journal of Energy Storage, 2016. 8: p. 142-159.
131. Pei, L., et al., Online peak power prediction based on a parameter and state estimator for lithium-ion batteries in electric vehicles. Energy, 2014. 66: p. 766-778.
132. Walder, G., et al. Adaptive state and parameter estimation of lithium-ion batteries based on a dual linear Kalman filter. in The Second International Conference on Technological Advances in Electrical, Electronics and Computer Engineering (TAEECE2014). 2014. The Society of Digital Information and Wireless Communication.
133. Mastali, M., et al., Battery state of the charge estimation using Kalman filtering. Journal of Power Sources, 2013. 239: p. 294-307.
134. Andre, D., et al., Advanced mathematical methods of SOC and SOH estimation for lithium-ion batteries. Journal of Power Sources, 2013. 224: p. 20-27.
135. Sepasi, S., R. Ghorbani, and B.Y. Liaw, Inline state of health estimation of lithium-ion batteries using state of charge calculation. Journal of Power Sources, 2015. 299: p. 246-254.
136. Hu, X., et al., Robustness analysis of State-of-Charge estimation methods for two types of Li-ion batteries. Journal of Power Sources, 2012. 217: p. 209-219.
137. Lagarias, J.C., et al., Convergence properties of the Nelder--Mead simplex method in low dimensions. SIAM Journal on optimization, 1998. 9(1): p. 112-147.
138. Wang, Q., et al., Correlation between the model accuracy and model-based SOC estimation. Electrochimica Acta, 2017.
110
139. Gibbs, B.P., Kalman Filtering, in Advanced Kalman Filtering, Least-Squares and Modeling. 2011, John Wiley & Sons, Inc. p. 289-342.
140. Mohinder, S.G., Introduction, in Kalman Filtering:Theory and Practice with MATLAB. 2015, Wiley-IEEE Press. p. 640.
141. Berecibar, M., et al., Online state of health estimation on NMC cells based on predictive analytics. Journal of Power Sources, 2016. 320: p. 239-250.
142. Welch, G. and G. Bishop, An introduction to the Kalman filter. 1995. 143. Kalman, R.E., A new approach to linear filtering and prediction problems. Journal of basic
Engineering, 1960. 82(1): p. 35-45. 144. Schneider, R. and C. Georgakis, How to not make the extended Kalman filter fail. Industrial &
Engineering Chemistry Research, 2013. 52(9): p. 3354-3362. 145. Almagbile, A., J. Wang, and W. Ding, Evaluating the performances of adaptive Kalman filter
methods in GPS/INS integration. Journal of Global Positioning Systems, 2010. 9(1): p. 33-40. 146. Plett, G.L., Sigma-point Kalman filtering for battery management systems of LiPB-based HEV
battery packs: Part 1: Introduction and state estimation. Journal of Power Sources, 2006. 161(2): p. 1356-1368.
147. He, Y., et al., A new model for State-of-Charge (SOC) estimation for high-power Li-ion batteries. Applied Energy, 2013. 101: p. 808-814.
148. He, H., Z. Liu, and Y. Hua, Adaptive Extended Kalman Filter Based Fault Detection and Isolation for a Lithium-Ion Battery Pack. Energy Procedia, 2015. 75: p. 1950-1955.
149. Wan, E.A. and A.T. Nelson, Dual Extended Kalman Filter Methods, in Kalman Filtering and Neural Networks. 2002, John Wiley & Sons, Inc. p. 123-173.
150. Plett, G.L., Sigma-point Kalman filtering for battery management systems of LiPB-based HEV battery packs: Part 2: Simultaneous state and parameter estimation. Journal of Power Sources, 2006. 161(2): p. 1369-1384.
151. Pudlák, P., The Complexity of Computations, in Logical Foundations of Mathematics and Computational Complexity: A Gentle Introduction. 2013, Springer International Publishing: Heidelberg. p. 365-493.
152. Sarasketa-Zabala, E., et al., Realistic lifetime prediction approach for Li-ion batteries. Applied Energy, 2016. 162: p. 839-852.
153. Xing, Y., et al., An ensemble model for predicting the remaining useful performance of lithium-ion batteries. Microelectronics Reliability, 2013. 53(6): p. 811-820.
154. Hussein, A.A. Experimental modeling and analysis of lithium-ion battery temperature dependence. in 2015 IEEE Applied Power Electronics Conference and Exposition (APEC). 2015.
111
7 APPENDICES
The following appendices contain the MATLAB code as used in the thesis to obtain the presented
results. For these algorithms to be used, some data needs to be loaded. The battery data set from the
NASA Ames Prognostics Center of Excellence called “Randomized Battery Usage Data Set” was used
throughout this work and is available here: https://ti.arc.nasa.gov/tech/dash/pcoe/prognostic-data-
repository/.
APPENDIX 1: CODE FOR PLOTTING AND ORGANIZATION OF BATTERY DIAGNOSTIC DATA This code was mostly adapted from the MATLAB file provided with the battery data. Capacity is
measured for all RW phases. These values were saved as “capacity” for later use in Kalman filter codes.
The structure “RWDis” containing all the indices of RW steps was also saved for later use.
close all; clear all; % Change this data file name below to plot different files load RW9
%% First we show a sample plot of the low current discharge data
steps = data.step; % save steps array to new variable
% search for instances of low current discharge step inds = []; % initialize index array for(i=1:length(steps)) if(strcmp(data.step(i).comment,'low current discharge at 0.04A')) inds = [inds, i]; end end
% plot data from first low current discharge step RT = steps(inds(1)).relativeTime/3600; % relative time in hours V = steps(inds(1)).voltage; % voltage I = steps(inds(1)).current; % current
figure % plot voltage profile plot(RT,V) title('Voltage for low current discharge') xlabel('Time (h)'); ylabel('Voltage (V)');
figure % plot current profile plot(RT,I) title('Current for low current discharge') ylim([0,1]) xlabel('Time (h)'); ylabel('Current (A)');
%% Next, the constant load profiles that are run after every 50 random walk % discharge cycles are plotted using
% initialize plot figure, hold on xlim([-.1,2.5]) ylim([3,4.25]) xlabel('Time (h)'); ylabel('Voltage (V)');
112
title('Reference Discharge Profiles')
for i = 1:length(steps) % search through the array of step structures if strcmp(steps(i).comment,'reference discharge') RT = steps(i).relativeTime/3600; % relative time in hours V = steps(i).voltage; % voltage plot(RT,V,'k') end end
%% We can benchmark the battery’s capacity by integrating current over the % reference cycles. The next plot shows this capacity measurement vs date.
counter = 0; % initialize counters
for i = 1:length(steps) % search through the array of step structures if strcmp(steps(i).comment,'reference discharge') counter = counter+1;
% save the date of this reference discharge step date(counter) = datenum(steps(i).date);
% save the battery capacity measured by integrating current capacity(counter) =
trapz(steps(i).relativeTime/3600,data.step(i).current); end end
% plot the data figure plot(date,capacity,'o') datetick('x','mmmyyyy') xlabel('Date'); ylabel('Capacity (Ah)'); title('Degradation of Measured Capacity')
%% The next code snippet shows how the resting periods after each reference % discharge can be grouped with the corresponding reference discharge cycle % and added to the plot above.
% initialize Reference Discharge Struct that will hold indexes of reference % discharge steps and the rest steps that come imediately before and after RefDis.repition{1}.indexes = 0;
% initialize loop parameters counter = 0; for i = 1:length(data.step) if strcmp(steps(i).comment,'reference discharge') counter = counter+1; if strcmp(steps(i-1).comment,'rest post reference charge') RefDis.repition{counter}.indexes = [i-1, i]; else RefDis.repition{counter}.indexes = i; end end
if strcmp(steps(i).comment,'rest post reference discharge') RefDis.repition{counter}.indexes =
[RefDis.repition{counter}.indexes, i];
113
end end
% now plot each of the grouped reference discharge cycles figure, hold on for i = 1:length(RefDis.repition) % stitch together all of the substeps identified in the repitition % field step = steps(RefDis.repition{i}.indexes(1));
% if the first index is a rest period, then make it end at relativeTime
== 0 if(strcmp(step.comment,'rest post reference charge')) step.relativeTime = step.relativeTime - step.relativeTime(end); end
for j = 2:length(RefDis.repition{i}.indexes) step.relativeTime = [step.relativeTime
step.relativeTime(end)+steps(RefDis.repition{i}.indexes(j)).relativeTime]; step.voltage = [step.voltage
steps(RefDis.repition{i}.indexes(j)).voltage]; step.current = [step.current
steps(RefDis.repition{i}.indexes(j)).current]; step.temperature = [step.temperature
steps(RefDis.repition{i}.indexes(j)).temperature]; end
plot(step.relativeTime/3600,step.voltage,'k') title('Reference discharge profiles and post-discharge rest periods') xlabel('Time (h)') ylabel('Voltage (V)') end
%% Example Plots of Pulsed Load Charging and Discharging Cycles
PulsedDis.repition{1}.indexes = 0; counter1 = 0; counter2 = 0; % find Pulsed Load discharge characterization cycles for i = 2:length(data.step) if (strcmp(data.step(i).comment,'pulsed load (discharge)')... || strcmp(data.step(i).comment,'pulsed load (rest)')) % if just starting the pulsed load cycle if (strcmp(data.step(i).comment,'pulsed load (rest)') &&
(~strcmp(data.step(i-1).comment,'pulsed load (discharge)'))) counter1 = 1; counter2 = counter2+1; PulsedDis.repition{counter2}.indexes(counter1) = i; else counter1 = counter1+1; PulsedDis.repition{counter2}.indexes(counter1) = i; end end if(strcmp(data.step(i).comment,'rest post pulsed load or charge')) % check to see if this preceeds a pulsed load (discharge) if(strcmp(data.step(i-1).comment,'pulsed load (discharge)')) counter1 = counter1+1; PulsedDis.repition{counter2}.indexes(counter1) = i; end end
114
end
% now plot each of the grouped pulsed load discharge cycles figure, hold on for i = 1:length(PulsedDis.repition) % stitch together all of the substeps identified in the repitition % field step = steps(PulsedDis.repition{i}.indexes(1));
for j = 2:length(PulsedDis.repition{i}.indexes) step.relativeTime = [step.relativeTime
step.relativeTime(end)+steps(PulsedDis.repition{i}.indexes(j)).relativeTime
]; step.voltage = [step.voltage
steps(PulsedDis.repition{i}.indexes(j)).voltage]; step.current = [step.current
steps(PulsedDis.repition{i}.indexes(j)).current]; step.temperature = [step.temperature
steps(PulsedDis.repition{i}.indexes(j)).temperature]; end
plot(step.relativeTime/3600,step.voltage,'k') title('Pulsed Discharge') xlabel('Time (h)') ylabel('Voltage (V)') end
%% Identify all random walk steps and plot current, voltage and temperature
for chosen steps and phase
RWDis.repition{1}.indexes = 0; counter1 = 0; counter2 = 0; lastStep = -1; % find indexes of all RW steps for i = 1:length(steps) if (strcmp(steps(i).comment,'discharge (random walk)')||... strcmp(steps(i).comment,'rest (random walk)')||... strcmp(steps(i).comment,'charge (random walk)')) % if there is a break in the step sequence then separate sequence % into new list if (i-lastStep > 1) counter1 = 1; counter2 = counter2+1; RWDis.repition{counter2}.indexes(counter1) = i; else counter1 = counter1+1; RWDis.repition{counter2}.indexes(counter1) = i; end lastStep = i; end end
% identify index sequence for first 50 RW steps from first RW phase RWsteps=50; RWphase=1; inds =
RWDis.repition{RWphase}.indexes(1):RWDis.repition{RWphase}.indexes(RWsteps*
2);
115
% stich all of the indexed steps together step = steps(inds(1)); for i = 2:length(inds) step.relativeTime = [step.relativeTime
step.relativeTime(end)+steps(inds(i)).relativeTime]; step.voltage = [step.voltage steps(inds(i)).voltage]; step.current = [step.current steps(inds(i)).current]; step.temperature = [step.temperature steps(inds(i)).temperature]; end
% plot Voltage figure subplot(3,1,1) plot(step.relativeTime/3600,step.voltage,'k') title(sprintf('First 50 RW steps',i)) xlabel('Time (h)') ylabel('Voltage (V)')
% plot Current subplot(3,1,2), hold on plot(step.relativeTime/3600,step.current,'k') xlabel('Time (h)') ylabel('Current (A)')
% plot temperature subplot(3,1,3), hold on plot(step.relativeTime/3600,step.temperature,'k') ylim([15, 40]) xlabel('Time (h)') ylabel('Temperature (C)')
APPENDIX 2: CODE FOR EXTENDED KALMAN FILTER close all, clear all
load RW9 % loading battery data file load RWDis % structure containing indeces of RW phases and steps in battery
data load capacity % vector containing all capacities for the battery in question
measured during reference discharge tests steps = data.step; % save steps array to new variable %% Data selection RWsteps=100; % number of RW steps from start of phase phase=1; % RW phase inds2 =
RWDis.repition{phase}.indexes(1):RWDis.repition{phase}.indexes(RWsteps*2); %% Stich all of the indexed steps together step = steps(inds2(1)); for i = 2:length(inds2) step.relativeTime = [step.relativeTime
step.relativeTime(end)+steps(inds2(i)).relativeTime]; % time of data points
(s) step.voltage = [step.voltage steps(inds2(i)).voltage]; % cell voltage (V) step.current = [step.current steps(inds2(i)).current]; % current (A) step.temperature = [step.temperature steps(inds2(i)).temperature]; %
temperature (deg. C) end %% Get cell capacity for RW phase from reference discharge tests step.capacity=(capacity(phase*2)+capacity(phase*2 - 1))/2; % average of the
two measurements before each RW phase
116
%% Coulomb counting for SOC reference for i = 1:length(step.current) if i==1 dt=step.relativeTime(i); % length of time step (s) SOC_0=1; % starting SOC value, representing a fully charged battery else dt=step.relativeTime(i)-step.relativeTime(i-1); current=(step.current(i)+step.current(i-1))/2; % average current
between timesteps SOC_0=SOC(i-1); end SOC(i)=SOC_0-(current*dt/(step.capacity*3600)); % Coulomb counting.
Conversion factor 3600 to convert capacities to unit A*s from A*h end %% Kalman initialization and internal parameters SOC_0=0.9; % starting guess of SOC V1_0=0; % starting guess of transient voltage (V1) x_hat_0=[V1_0; SOC_0]; %Starting state vector x covx_p=[1e-3 0;0 1e-2]; % initial state covariance matrix covw=[1e-6 0;0 1e-10]; % process noise covariance matrix covv=5e-3; % measurement noise covariance (matrix) %% Adding artifial noise to signal step.voltage=step.voltage.*(1 + randn(1,length(step.voltage)) * 5e-3 ); %
adding artificial random noise to measurement signal %% ECM parameters from pulsed discharge test R0=0.074; % Ohm R1=0.045; % Ohm C1=815; % Farad %% Extended Kalman filtering for i=1:length(step.relativeTime) % filtering over all selected time steps %% Time update if i==1 % time step dt=step.relativeTime(i); else dt=step.relativeTime(i)-step.relativeTime(i-1); end A=[ exp(-dt/(R1*C1)) 0 ; 0 1 ]; % calculating A matrix B=[ R1 * (1-exp(-dt/(R1*C1))) ; -dt/(step.capacity*3600) ]; % calculating
B matrix if i==1 % state estimate time update x_hat_minus=A*x_hat_0 + B*step.current(i); else x_hat_minus=A*x_hat_plus(:,i-1) + B*step.current(i-1); end covx_m=A*covx_p*A' + covw; % state covariance matrix time update [OCV, dOCVdSOC]= OCV_SOC(x_hat_minus(2)); % function for the OCV-SOC
relationship, also outputs the derivative of this function. C(i,:)=[-1, dOCVdSOC]; % calculating C matrix Yest(i)=OCV - R0*step.current(i) - x_hat_minus(1); % output equation %% Measurement update err(i)=step.voltage(i)-Yest(i); % error of predicted cell voltage L(:,i)=covx_m*C(i,:)'*( (C(i,:)*covx_m*C(i,:)' + covv )^-1); %
calculating Kalman gain matrix x_hat_plus(:,i)=x_hat_minus + L(:,i)*err(i); % state estimate measurement
update covx_p=( eye(2) - L(:,i)*C(i,:) )*covx_m; % state covariance matrix
measurement update end %% Plotting results figure % plot of Kalman SOC estimation in comparison to Coulomb counting SOC
117
plot(step.relativeTime/3600,SOC,step.relativeTime/3600,x_hat_plus(2,:),'Lin
eWidth',2) legend('Coulomb counting SOC','EKF SOC'); set(gca,'FontSize',22) title('EKF SOC vs. Coulomb counting SOC') ylim([0, 1]) xlabel('Time (h)') ylabel('SOC')
APPENDIX 3: CODE FOR ADAPTIVE EXTENDED KALMAN FILTER close all, clear all, clc load RW9 % loading battery data file load RWDis % structure containing indeces of RW phases and steps in battery
data load capacity % vector containing all capacities for the battery in question
measured during reference discharge tests steps = data.step; % save steps array to new variable %% Data selection RWsteps=100; % number of RW steps from start of phase phase=1; % RW phase inds2 =
RWDis.repition{phase}.indexes(1):RWDis.repition{phase}.indexes(RWsteps*2); %% Stich all of the indexed steps together step = steps(inds2(1)); for i = 2:length(inds2) step.relativeTime = [step.relativeTime
step.relativeTime(end)+steps(inds2(i)).relativeTime]; % time of data points
(s) step.voltage = [step.voltage steps(inds2(i)).voltage]; % cell voltage (V) step.current = [step.current steps(inds2(i)).current]; % current (A) step.temperature = [step.temperature steps(inds2(i)).temperature]; %
temperature (deg. C) end %% Get cell capacity for RW phase from reference discharge tests step.capacity=(capacity(phase*2)+capacity(phase*2 - 1))/2; % average of the
two measurements before each RW phase %% Coulomb counting for SOC reference for i = 1:length(step.current) if i==1 dt=step.relativeTime(i); % length of time step (s) SOC_0=1; % starting SOC value, representing a fully charged battery else dt=step.relativeTime(i)-step.relativeTime(i-1); current=(step.current(i)+step.current(i-1))/2; % average current
between timesteps SOC_0=SOC(i-1); end SOC(i)=SOC_0-(current*dt/(step.capacity*3600)); % Coulomb counting.
Conversion factor 3600 to convert capacities to unit A*s from A*h end
%% Kalman initialization. %% Kalman initialization and internal parameters SOC_0=0.9; % starting guess of SOC V1_0=0; % starting guess of transient voltage (V1) x_hat_0=[V1_0; SOC_0]; %Starting state vector x covx_p=[1e-3 0;0 1e-2]; % initial state covariance matrix M=10; % moving estimation window size for AEKF Q_0=[1e-2 0 ; 0 1e-2]; % starting process noise covariance matrix
118
R_0=1e-2; % starting measurement noise covariance %% Adding artifial noise to signal step.voltage=step.voltage.*(1 + randn(1,length(step.voltage)) * 5e-3 ); %
adding artificial random noise to measurement signal %% ECM parameters from pulsed discharge test R0=0.074; % Ohm R1=0.045; % Ohm C1=815; % Farad %% Kalman filtering for i=1:length(step.voltage) if i==1 % time step dt=step.relativeTime(i); else dt=step.relativeTime(i)-step.relativeTime(i-1); end A=[ exp(-dt/(R1*C1)) 0 ; 0 1 ]; B=[ R1 * (1-exp(-dt/(R1*C1))) ; -dt/(step.capacity*3600) ];
if i==1 x_hat_minus=A*x_hat_0 + B*I(i); % % time update state estimate covx_m=A*covx_p*A' + Q_0; % time update state covariance matrix else x_hat_minus=A*x_hat_plus(:,i-1) + B*I(i-1); covx_m=A*covx_p*A' + Q(:,:,i-1); end [OCV, dOCVdSOC]= OCV_SOC(x_hat_minus(2)); % function for the OCV-SOC
relationship, also outputs the derivative of this function C(i,:)=[-1, dOCVdSOC ]; % calculating C matrix Yest(i)=OCV - R0*I(i) - x_hat_minus(1); % output equation err(i)=Y(i)-Yest(i); % error of predicted cell voltage %% Updating measurement noise covariance if i<=M R(i)=R_0; else H(i)=(1/M)*sum(err(i-M+1:i).^2); R(i)=H(i) + C(i,:)*covx_m*C(i,:)'; end %% Kalman gain and measurement updates to x and its covariance if i==1 L(:,i)=( covx_m*C(i,:)' )*( (C(i,:)*covx_m*C(i,:)' + R_0 )^-1); %
Kalman gain matrix x_hat_plus(:,i)=x_hat_minus +L(:,i)*err(i); % measurement state
update covx_p=( eye(2) - L(:,i)*C(i,:) )*covx_m; else L(:,i)=( covx_m*C(i,:)' )*( (C(i,:)*covx_m*C(i,:)' + R(i) )^-1); %
Kalman gain matrix x_hat_plus(:,i)=x_hat_minus +L(:,i)*err(i); % measurement state
update covx_p=( eye(2) - L(:,i)*C(i,:) )*covx_m; end %% Updating process noise covariance if i<=M Q(:,:,i)=Q_0; else Q(:,:,i)=L(:,i)*H(i)*L(:,i)'; end
end %% plotting results figure % plot of Kalman SOC estimation in comparison to Coulomb counting
119
plot(step.relativeTime/3600,SOC,step.relativeTime/3600,x_hat_plus(2,:),'Lin
eWidth',2) legend('CC SOC','AEKF SOC'); set(gca,'FontSize',22) title('Kalman SOC') ylim([0, 1]) xlabel('Time (h)') ylabel('SOC')
APPENDIX 4: CODE FOR DUAL EXTENDED KALMAN FILTER close all, clear all load RW9 % loading battery data file load RWDis % structure containing indeces of RW phases and steps in battery
data load capacity % vector containing all capacities for the battery in question
measured during reference discharge tests steps = data.step; % save steps array to new variable %% Data selection RWsteps=100; % number of RW steps from start of phase phase=1; % RW phase inds2 =
RWDis.repition{phase}.indexes(1):RWDis.repition{phase}.indexes(RWsteps*2); %% stich all of the indexed steps together step = steps(inds2(1)); for i = 2:length(inds2) step.relativeTime = [step.relativeTime
step.relativeTime(end)+steps(inds2(i)).relativeTime]; % time of data points
(s) step.voltage = [step.voltage steps(inds2(i)).voltage]; % cell voltage (V) step.current = [step.current steps(inds2(i)).current]; % current (A) step.temperature = [step.temperature steps(inds2(i)).temperature]; %
temperature (deg. C) end %% get cell capacity for RW phase from reference discharge tests step.capacity=(capacity(phase*2)+capacity(phase*2 - 1))/2; % used cell
capacity is average of the two measurements before each RW phase %% Coulomb counting for SOC reference for i = 1:length(step.current) if i==1 dt=step.relativeTime(i); I=step.current(i); SOC_0=1; % starting value else dt=step.relativeTime(i)-step.relativeTime(i-1); I=(step.current(i)+step.current(i-1))/2; % average current between
timesteps SOC_0=SOC(i-1); end SOC(i)=SOC_0-(I*dt/(step.capacity*3600)); % adjusted capacity for
temperature according to Farmann, 2015: 7.7% higher capacity per 30 deg C,
average temp. of capacity measurement: 24.96 deg C, average temp. of 50 first
RW cycles: 31.94 deg C. end %% Dual Kalman filter initialization %% State filter internal paramters SOC_0=0.999; % starting guess of SOC V1_0=0; % starting guess of transient voltage (V1) x_hat_0=[ V1_0 ; SOC_0 ]; % initial state vector covx_plus=[1e-5 0;0 1e-10]; % initial state covariance matrix for state
filter
120
sigmav=5e-3; % measurement noise covariance (matrix) for state filter covw=[ 1e-8 0 ; 0 1e-10 ]; % process noise covariance matrix state filter %% weight filter internal parameters cap_0=1.9 % starting guess of cell capacity R0_0=0.074; % starting guess of R0 R1_0=0.045; % starting guess of R1 covohm_plus=[1e-9 0 0 ; 0 1e-9 0 ; 0 0 1e-4]; % initial state covariance
matrix for weight filter sigmae=1e-2 ; % measurement noise covariance (matrix) for weight filter covr=[ 1e-11 0 0 ; 0 1e-11 0 ; 0 0 1e-9 ]; % process noise covariance
matrix for weight filter %% Adding artifial noise to signal step.voltage=step.voltage.*(1 + randn(1,length(step.voltage)) * 5e-3 );
% adding artificial random noise to measurement signal %% starting guesses for cell parameters C1=815.1; % Farad ohm_hat_0=[ R0_0 ; R1_0 ; cap_0]; % weight vector %% Dual extended Kalman filtering for i=1:length(step.relativeTime) % filtering over all selected time steps %% time update weight filter then state filter if i==1 % time step dt(i)=step.relativeTime(i); else dt(i)=step.relativeTime(i)-step.relativeTime(i-1); end if i==1 ohm_hat_minus=ohm_hat_0; % weight filter time update A(:,:,i)=[ exp(-dt(i)/(ohm_hat_0(2)*C1)) 0 ; 0 1]; B=[ ohm_hat_0(2) * (1-exp(-dt(i)/(ohm_hat_0(2)*C1))) ; -
dt(i)/(ohm_hat_0(3)*3600)]; x_hat_minus=A(:,:,i)*x_hat_0 + B*I(i); % state filter time update else ohm_hat_minus=ohm_hat_plus(:,i-1); A(:,:,i)=[ exp(-dt(i)/(ohm_hat_minus(2)*C1)) 0 ; 0 1]; B=[ ohm_hat_minus(2) * (1-exp(-dt(i)/(ohm_hat_minus(2)*C1))) ; -
dt(i)/(ohm_hat_minus(3)*3600)]; x_hat_minus=A(:,:,i)*x_hat_plus(:,i-1) + B*I(i-1); end covohm_minus=covohm_plus + covr; % time update weight filter covariance
matrix covx_minus=A(:,:,i)*covx_plus*A(:,:,i)' + covw; % time update state
filter covariance matrix [OCV, dOCVdSOC]= OCV_SOC(x_hat_minus(2)); % function for the OCV-SOC
relationship, also outputs the derivative of this curve. %% calculating C matrices for both filters C(i,:)=[-1,dOCVdSOC ]; %derivative of OCV-SOC polynomial if i==1 % calculation of C matrix for weight filter requires recursive
procedure C_ohm(i,:)=[0 0 0]; % initializing values F(:,:,i)=[0 0 0 ; 0 0 0]; % dx_plus_k-1/dohm E(:,:,i)=[0 0 0 ; 0 0 0]; % dx_minus_k/dohm else %% partial derivatives dSOCdC(i)=I(i)*dt(i)/( 3600*( ohm_hat_minus(3) )^2 ); %dSOC/dC dgdC(i)=0; %dg/dC dxdR1_plus1(i)= x_hat_plus(1,i-1)* dt(i) * exp( -
dt(i)/(ohm_hat_minus(2)*C1)) / (C1*( ohm_hat_minus(2) )^2); dxdR1_plus2(i)=I(i-1)*( 1 - exp( -dt(i)/(ohm_hat_minus(2)*C1) ) -
dt(i)*exp( -dt(i)/(ohm_hat_minus(2)*C1) )/( ohm_hat_minus(2)*C1 ) ); dxdR1_plus(i)=dxdR1_plus1(i)+dxdR1_plus2(i); %dV1/dR1 at
x_hat_plus_k-1
121
dxdR1_minus1(i)= x_hat_minus(1)* dt(i) * exp( -
dt(i)/(ohm_hat_minus(2)*C1)) / (C1*( ohm_hat_minus(2) )^2); dxdR1_minus2(i)=I(i)*( 1 - exp( -dt(i)/(ohm_hat_minus(2)*C1) ) -
dt(i)*exp( -dt(i)/(ohm_hat_minus(2)*C1) )/( ohm_hat_minus(2)*C1 ) ); dxdR1_minus(i)=dxdR1_minus1(i)+dxdR1_minus2(i); %dV1/dR1 at current
x_hat_minus %% full Jacobians dgdohm=[-I(i), -dxdR1_minus(i), dgdC(i)]; % [dg/dR0 dg/dR1 dg/dC] dfdohm=[0 , dxdR1_plus(i) , 0 ; 0 , 0 , dSOCdC(i-1)]; % [dV1/dR0
dV1/dR1 dV1/dC; dSOC/dR0 dSOC/dR1 dSOC/dC] %% recursive derivatives F(:,:,i)=E(:,:,i-1) - L(:,i-1)*C_ohm(i-1,:); E(:,:,i)=dfdohm + A(:,:,i)*F(:,:,i); C_ohm(i,:)=dgdohm + C(i,:)*E(:,:,i); % dg/dohm end %% measurement update for state filter L(:,i)=covx_minus*C(i,:)' * ( (C(i,:) * covx_minus * C(i,:)' + sigmav )^-
1 ); % Kalman gain matrix for state filter Yest(i)=OCV - ohm_hat_minus(1)*I(i) - x_hat_minus(1); % output equation err(i)=Y(i)-Yest(i); % calculation of error of cell voltage prediction x_hat_plus(:,i)=x_hat_minus + L(:,i)*err(i); % state estimate measurement
update covx_plus=( eye(2) - L(:,i)*C(i,:) )*covx_minus; % state filter
covariance matrix measurement update %% measurement update for weight filter L_ohm(:,i)= covohm_minus *C_ohm(i,:)' * ( ( C_ohm(i,:) * covohm_minus *
C_ohm(i,:)' + sigmae )^-1 ); % Kalman gain matrix for weight filter ohm_hat_plus(:,i)=ohm_hat_minus + L_ohm(:,i)*err(i); % state estimate
measurement update covohm_plus=( eye(3) - L_ohm(:,i)*C_ohm(i,:) )*covohm_minus; % weight
filter covariance matrix measurement update end %% plotting of results figure % plot of Kalman SOC estimation in comparison to Coulomb counting plot(step.relativeTime/3600,SOC,step.relativeTime/3600,x_hat_plus(2,:),'Lin
ewidth',2) legend('CC SOC','Kalman SOC'); title('Kalman SOC') set(gca,'FontSize',22) ylim([0, 1]) xlim([0, max(step.relativeTime/3600)]) xlabel('Time (h)') ylabel('SOC')
figure % plot of estimated capacity plot(step.relativeTime/3600,ohm_hat_plus(3,:),'Linewidth',2) title('Estimated capacity') set(gca,'FontSize',22) legend('Capacity (Ah)') ylim([1.5, 2.3]) xlabel('Time (h)') ylabel('Capacity (Ah)')
figure % plot of R0 and R1 plot(step.relativeTime/3600,ohm_hat_plus(1,:),step.relativeTime/3600,ohm_ha
t_plus(2,:),'Linewidth',2) set(gca,'FontSize',22) title('Estimated internal resistances') legend('R_0','R_1') set(gca,'FontSize',22) xlabel('Time (h)')
122
ylabel('Resistance (Ohm)')
APPENDIX 5: CODE FOR ENHANCED STATE VECTOR EXTENDED KALMAN FILTER close all, clear all load RW9 % loading battery data file load RWDis % structure containing indeces of RW phases and steps in battery
data load capacity % vector containing all capacities for the battery in question
measured during reference discharge tests steps = data.step; % save steps array to new variable %% Data selection RWsteps=100; % number of RW steps from start of phase phase=1; % RW phase inds2 =
RWDis.repition{phase}.indexes(1):RWDis.repition{phase}.indexes(RWsteps*2); %% Stich all of the indexed steps together step = steps(inds2(1)); for i = 2:length(inds2) step.relativeTime = [step.relativeTime
step.relativeTime(end)+steps(inds2(i)).relativeTime]; % time of data points
(s) step.voltage = [step.voltage steps(inds2(i)).voltage]; % cell voltage (V) step.current = [step.current steps(inds2(i)).current]; % current (A) step.temperature = [step.temperature steps(inds2(i)).temperature]; %
temperature (deg. C) end %% Get cell capacity for RW phase from reference discharge tests step.capacity=(capacity(phase*2)+capacity(phase*2 - 1))/2; % average of the
two measurements before each RW phase %% Coulomb counting for SOC reference for i = 1:length(step.current) if i==1 dt=step.relativeTime(i); % length of time step (s) SOC_0=1; % starting SOC value, representing a fully charged battery else dt=step.relativeTime(i)-step.relativeTime(i-1); current=(step.current(i)+step.current(i-1))/2; % average current
between timesteps SOC_0=SOC(i-1); end SOC(i)=SOC_0-(current*dt/(step.capacity*3600)); % Coulomb counting.
Conversion factor 3600 to convert capacities to unit A*s from A*h end %% Kalman initialization and internal parameters SOC_0=0.999; % starting guess of SOC V1_0=0; % starting guess of transient voltage (V1) (V) cap_0=1.9; % starting guess of cell capacity (Ah) x_hat_0=[V1_0; SOC_0; 1/cap_0]; %Starting state vector x covx_p=[1e-4 0 0;0 1e-10 0;0 0 1e-4]; % initial state covariance matrix covw=[1e-4 0 0 ; 0 1e-10 0 ; 0 0 1e-11]; % process noise covariance matrix covv=5e-3; % measurement noise covariance (matrix) %% ECM parameters from pulsed discharge test R0=0.074; % Ohm R1=0.045; % Ohm C1=815.1; % Farad %% Kalman filter with enhanced state vector for i=1:length(step.voltage) % filtering over all selected time steps %% Time update if i==1 dt=step.relativeTime(i); % time step (s)
123
A=[ exp(-dt/(R1*C1)) 0 0 ; 0 1 -dt*step.current(i)/3600 ; 0 0 1 ]; B=[ R1 * (1-exp(-dt/(R1*C1))) ; 0 ; 0 ]; x_hat_minus=A*x_hat_0 + B*step.current(i); % state estimate time
update else dt=step.relativeTime(i)-step.relativeTime(i-1); A=[ exp(-dt/(R1*C1)) 0 0 ; 0 1 -dt*I(i-1)/3600 ; 0 0 1 ]; B=[ R1 * (1-exp(-dt/(R1*C1))) ; 0 ; 0 ]; x_hat_minus=A*x_hat_plus(:,i-1) + B*step.current(i-1); end covx_m=A*covx_p*A' + covw; % state covariance matrix time update [OCV, dOCVdSOC]=OCV_SOC(x_hat_minus(2)); % function for the OCV-SOC
relationship, also outputs the derivative of this curve %% Calculating the C matrix dSOCdC(i)=-I(i)*dt/3600; %dSOC/d(1/C)) as x(3)=1/C. dgdC(i)=dOCVdSOC*dSOCdC(i); C(i,:)=[-1, dOCVdSOC, dgdC(i) ]; %C=dg/dx Yest(i)=OCV - R0*I(i) - x_hat_minus(1); % output equation err(i)=Y(i)-Yest(i); % error of predicted cell voltage %% Measurement update L(:,i)=covx_m*C(i,:)'*( (C(i,:)*covx_m*C(i,:)' + covv )^-1); % Kalman
gain x_hat_plus(:,i)=x_hat_minus +L(:,i)*err(i); % state estimate measurement
update covx_p=( eye(3) - L(:,i)*C(i,:) )*covx_m; % state covariance matrix
measurement update end %% plotting results figure % plot of Kalman SOC estimation in comparison to Coulomb counting plot(step.relativeTime/3600,SOC,step.relativeTime/3600,x_hat_plus(2,:),'Lin
eWidth',2) legend('CC SOC','Kalman SOC'); set(gca,'FontSize',22) title('Kalman SOC') ylim([0, 1]) xlabel('Time (h)') ylabel('SOC')
figure % plot of the estimated capacity plot(step.relativeTime/3600,1./x_hat_plus(3,:),'LineWidth',2) set(gca,'FontSize',22) ylim([1.5 2.3]) xlabel('Time (h)'); ylabel('Capacity (Ah)');
APPENDIX 6: CODE FOR SINGLE WEIGHT DUAL EXTENDED KALMAN FILTER close all, clear all, clc load RW9 % loading battery data file load RWDis % structure containing indeces of RW phases and steps in battery
data load capacity % vector containing all capacities for the battery in question
measured during reference discharge tests steps = data.step; % save steps array to new variable %% Data selection RWsteps=100; % number of RW steps from start of phase phase=1; % RW phase inds2 =
RWDis.repition{phase}.indexes(1):RWDis.repition{phase}.indexes(RWsteps*2); %% Stich all of the indexed steps together step = steps(inds2(1));
124
for i = 2:length(inds2) step.relativeTime = [step.relativeTime
step.relativeTime(end)+steps(inds2(i)).relativeTime]; % time of data points
(s) step.voltage = [step.voltage steps(inds2(i)).voltage]; % cell voltage (V) step.current = [step.current steps(inds2(i)).current]; % current (A) step.temperature = [step.temperature steps(inds2(i)).temperature]; %
temperature (deg. C) end %% Get cell capacity for RW phase from reference discharge tests step.capacity=(capacity(phase*2)+capacity(phase*2 - 1))/2; % average of the
two measurements before each RW phase %% Coulomb counting for SOC reference for i = 1:length(step.current) if i==1 dt=step.relativeTime(i); % length of time step (s) SOC_0=1; % starting SOC value, representing a fully charged battery else dt=step.relativeTime(i)-step.relativeTime(i-1); current=(step.current(i)+step.current(i-1))/2; % average current
between timesteps SOC_0=SOC(i-1); end SOC(i)=SOC_0-(current*dt/(step.capacity*3600)); % Coulomb counting.
Conversion factor 3600 to convert capacities to unit A*s from A*h end %% Kalman initialization and internal parameters %% State filter internal parameters SOC_0=0.999; % starting guess of SOC V1_0=0; % starting guesof transient voltage (V) x_hat_0=[ V1_0 ; SOC_0 ]; % starting state vector covx_plus=[1e-6 0;0 1e-8]; % guess initial covariance matrix for state
filter sigmav=1e-8; % measurement noise covariance matrix for state filter covw=[ 1e-8 0 ; 0 1e-12 ]; % state filter process noise covariance matrix %% Weight filter internal parameters covohm_plus=1e-3; % guess initial covariance matrix for weight filter sigmae=1e-4; % measurement noise covariance (matrix) for weight filter covr=1e-11 ; % weight filter process noise covariance (matrix) %% Starting guesses/values of cell parameters and ECM parameters cap_0=1.9; % initial guess of cell capacity (Ah) R0=0.074; % Ohm R1=0.045; % Ohm C1=815; % Farad ohm_hat_0=cap_0; % weight filter state (vector) %% Single weight dual extended Kalman filtering % "minus" indicates estimates after time update and "plus" estimates after % measurement update for i=1:length(step.voltage) %% time update of both filters if i==1 % time step dt(i)=step.relativeTime(i); ohm_hat_minus=ohm_hat_0; % time update weight filter A=[ exp(-dt(i)/(R1*C1)) 0 ; 0 1]; B=[ R1 * (1-exp(-dt(i)/(R1*C1))) ; -dt(i)/(ohm_hat_minus*3600)]; x_hat_minus=A*x_hat_0 + B*I(i); % time update state filter else dt(i)=step.relativeTime(i)-step.relativeTime(i-1); ohm_hat_minus=ohm_hat_plus(i-1); A=[ exp(-dt(i-1)/(R1*C1)) 0 ; 0 1]; B=[ R1 * (1-exp(-dt(i-1)/(R1*C1))) ; -dt(i-1)/(ohm_hat_minus*3600)];
125
x_hat_minus=A*x_hat_plus(:,i-1) + B*I(i-1); end covohm_minus=covohm_plus + covr; % time update weight filter covariance
matrix covx_minus=A*covx_plus*A' + covw; % time update state covariance matrix [OCV, dOCVdSOC]=OCV_SOC(x_hat_minus(2)); % function for the OCV-SOC
relationship, also outputs the derivative of this curve. C(i,:)=[-1, dOCVdSOC ]; % C_x=dg/dx %% Measurement update for state filter as well as calculation of
error L(:,i)=covx_minus*C(i,:)' * ( (C(i,:) * covx_minus * C(i,:)' +
sigmav )^-1 ); % Kalman gain matrix for state filter Yest(i)=OCV - R0*I(i) - x_hat_minus(1); % output equation err(i)=Y(i)-Yest(i); % error of predicted cell voltage x_hat_plus(:,i)=x_hat_minus + L(:,i)*err(i); % state filter
measurement update covx_plus=( eye(2) - L(:,i)*C(i,:) )*covx_minus; % state filter
covariance matrix measurement update %% C matrix for weight filter if i==1 % calculation of C matrix for capacity filter requires
recursive derivatives C_ohm(i,:)=0; % initializing values F(:,i)=[0 ; 0]; % dx_plus_k-1/dC E(:,i)=[0 ; 0]; % dx_minus_k/dC else %% Partial derivatives dSOCdC(i)=I(i)*dt(i)/(3600*(ohm_hat_minus)^2); % inner
derivative dgdC(i)=0; dfdC=[0 ; dSOCdC(i-1)]; %% Recursive derivatives F(:,i)=E(:,i-1) - L(:,i-1)*C_ohm(i-1); E(:,i)=dfdC + A*F(:,i); C_ohm(i)=dgdC(i) + C(i,:)*E(:,i); end %% Measurement update for weight filter L_ohm(i)= covohm_minus *C_ohm(i)' * ( ( C_ohm(i) * covohm_minus *
C_ohm(i)' + sigmae )^-1 ); % Kalman gain matrix for weight filter ohm_hat_plus(i)=ohm_hat_minus + L_ohm(:,i)*err(i); % updating weight
estimate post measurement covohm_plus=( 1 - L_ohm(i) * C_ohm(i) )*covohm_minus; % weight filter
covariance matrix measurement update end %% Plotting of results figure % plot of Kalman SOC estimation in comparison to Coulomb counting plot(step.relativeTime/3600,SOC,step.relativeTime/3600,x_hat_plus(2,:),'Lin
eWidth',2) legend('CC SOC','Kalman SOC'); title('Kalman SOC') set(gca,'FontSize',22) ylim([0, 1]) xlabel('Time (h)') ylabel('SOC')
figure % plot of estimated capacity plot(step.relativeTime/3600,ohm_hat_plus,'LineWidth',2) title('Estimated capacity') set(gca,'FontSize',22) ylim([1.8, 2.3]) legend('Capacity (Ah)') xlabel('Time (h)')
126
ylabel('Capacity (Ah)')